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_3x8c8__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_3x8c8__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 <= 3);
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 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
51 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56
57 do {
58 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
59 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
60 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
61 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
62 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
63 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
64 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
65 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
66 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
67 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
68 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
69 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
70 __m256i vacc1x01 = vacc0x01;
71 __m256i vacc1x23 = vacc0x23;
72 __m256i vacc1x45 = vacc0x45;
73 __m256i vacc1x67 = vacc0x67;
74 __m256i vacc2x01 = vacc0x01;
75 __m256i vacc2x23 = vacc0x23;
76 __m256i vacc2x45 = vacc0x45;
77 __m256i vacc2x67 = vacc0x67;
78 w = (const int32_t*) w + 8;
79
80 size_t k = 0;
81 const __m256i vb_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.kernel_zero_point);
82 while (k < kc) {
83 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
84 const __m256i vxa0 = _mm256_cvtepu8_epi16(va0);
85 a0 += 8;
86 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
87 const __m256i vxa1 = _mm256_cvtepu8_epi16(va1);
88 a1 += 8;
89 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
90 const __m256i vxa2 = _mm256_cvtepu8_epi16(va2);
91 a2 += 8;
92
93 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
94 const __m256i vxb01 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb01), vb_zero_point);
95
96 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
97 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
98 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
99 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 16));
100 const __m256i vxb23 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb23), vb_zero_point);
101
102 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
103 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
104 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
105 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 32));
106 const __m256i vxb45 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb45), vb_zero_point);
107
108 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
109 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
110 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
111 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 48));
112 const __m256i vxb67 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb67), vb_zero_point);
113
114 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
115 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
116 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
117
118 w = (const void*) ((const uint8_t*) w + 64);
119 k += 8 * sizeof(uint8_t);
120 }
121
122 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
123 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
124 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
125 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
126 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
127 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
128
129 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
130 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
131 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
132
133 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
134 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
135 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
136 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
137
138 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
139 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
140 __m256 vscaled2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
141
142 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
143 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale);
144 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale);
145 vscaled2x01234567 = _mm256_mul_ps(vscaled2x01234567, vscale);
146
147 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
148 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
149 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
150 vscaled2x01234567 = _mm256_min_ps(vscaled2x01234567, voutput_max_less_zero_point);
151
152 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
153 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
154 vacc2x01234567 = _mm256_cvtps_epi32(vscaled2x01234567);
155
156 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
157 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
158 __m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), voutput_zero_point);
159
160 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
161 vacc22x01234567 = _mm256_permute4x64_epi64(vacc22x01234567, _MM_SHUFFLE(3, 1, 2, 0));
162
163 __m256i vout = _mm256_packus_epi16(vacc01x01234567, vacc22x01234567);
164
165 vout = _mm256_max_epu8(vout, _mm256_load_si256((const __m256i*) params->fp32_avx2.output_min));
166
167 __m128i vout_lo = _mm256_castsi256_si128(vout);
168 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
169
170 if (nc >= 8) {
171 _mm_storel_epi64((__m128i*) c0, vout_lo);
172 _mm_storel_epi64((__m128i*) c1, vout_hi);
173 _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
174
175 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
176 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
177 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
178
179 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
180 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
181 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
182
183 nc -= 8;
184 } else {
185 if (nc & 4) {
186 _mm_storeu_si32(c0, vout_lo);
187 _mm_storeu_si32(c1, vout_hi);
188 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout_lo, 2));
189
190 c0 += 4;
191 c1 += 4;
192 c2 += 4;
193
194 vout_lo = _mm_srli_epi64(vout_lo, 32);
195 vout_hi = _mm_srli_epi64(vout_hi, 32);
196 }
197 if (nc & 2) {
198 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
199 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout_hi, 0));
200 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout_lo, 4));
201
202 c0 += 2;
203 c1 += 2;
204 c2 += 2;
205
206 vout_lo = _mm_srli_epi32(vout_lo, 16);
207 vout_hi = _mm_srli_epi32(vout_hi, 16);
208 }
209 if (nc & 1) {
210 *c0 = (uint8_t) _mm_extract_epi8(vout_lo, 0);
211 *c1 = (uint8_t) _mm_extract_epi8(vout_hi, 0);
212 *c2 = (uint8_t) _mm_extract_epi8(vout_lo, 8);
213 }
214
215 nc = 0;
216 }
217 } while (nc != 0);
218 }
219