Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-minmax-fp32-avx256skx.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_4x8c8__avx256skx(
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 <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * 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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 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);
61 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
62 const __m256i voutput_min = _mm256_set1_epi8(params->fp32_scalar.output_min);
63 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
64 // XNN_FORCE_REALIZATION(voutput_zero_point);
65 // XNN_FORCE_REALIZATION(voutput_min);
66
67 do {
68 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
69 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
70 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
71 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
72 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
73 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
74 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
75 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
76 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
77 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
78 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
79 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
80 __m256i vacc1x01 = vacc0x01;
81 __m256i vacc1x23 = vacc0x23;
82 __m256i vacc1x45 = vacc0x45;
83 __m256i vacc1x67 = vacc0x67;
84 __m256i vacc2x01 = vacc0x01;
85 __m256i vacc2x23 = vacc0x23;
86 __m256i vacc2x45 = vacc0x45;
87 __m256i vacc2x67 = vacc0x67;
88 __m256i vacc3x01 = vacc0x01;
89 __m256i vacc3x23 = vacc0x23;
90 __m256i vacc3x45 = vacc0x45;
91 __m256i vacc3x67 = vacc0x67;
92 w = (const int32_t*) w + 8;
93
94 size_t p = ks;
95 do {
96 const int8_t* restrict a0 = a[0];
97 if XNN_UNPREDICTABLE(a0 != zero) {
98 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
99 }
100 const int8_t* restrict a1 = a[1];
101 if XNN_UNPREDICTABLE(a1 != zero) {
102 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
103 }
104 const int8_t* restrict a2 = a[2];
105 if XNN_UNPREDICTABLE(a2 != zero) {
106 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
107 }
108 const int8_t* restrict a3 = a[3];
109 if XNN_UNPREDICTABLE(a3 != zero) {
110 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
111 }
112 a += 4;
113
114 size_t k = 0;
115 while (k < kc) {
116 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
117 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
118 a0 += 8;
119 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
120 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
121 a1 += 8;
122 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
123 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
124 a2 += 8;
125 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
126 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
127 a3 += 8;
128
129 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
130 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
131
132 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
133 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
134 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
135 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
136 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
137 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
138
139 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
140 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
141 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
142 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
143 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
144 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
145
146 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
147 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
148 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
149 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
150 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
151 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
152
153 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
154 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
155 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
156 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
157
158 w = (const void*) ((const int8_t*) w + 64);
159 k += 8 * sizeof(int8_t);
160 }
161 p -= 4 * sizeof(void*);
162 } while (p != 0);
163
164 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
165 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
166 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
167 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
168 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
169 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
170 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
171 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
172
173 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
174 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
175 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
176 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
177
178 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
179 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
180 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
181 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
182 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
183
184 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
185 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
186 __m256 vscaled2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
187 __m256 vscaled3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
188
189 const __m256 vscale01234567 = _mm256_load_ps(w);
190 w = (const void*) ((const float*) w + 8);
191 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale01234567);
192 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale01234567);
193 vscaled2x01234567 = _mm256_mul_ps(vscaled2x01234567, vscale01234567);
194 vscaled3x01234567 = _mm256_mul_ps(vscaled3x01234567, vscale01234567);
195
196 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
197 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
198 vscaled2x01234567 = _mm256_min_ps(vscaled2x01234567, voutput_max_less_zero_point);
199 vscaled3x01234567 = _mm256_min_ps(vscaled3x01234567, voutput_max_less_zero_point);
200
201 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
202 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
203 vacc2x01234567 = _mm256_cvtps_epi32(vscaled2x01234567);
204 vacc3x01234567 = _mm256_cvtps_epi32(vscaled3x01234567);
205
206 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
207 __m256i vacc23x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc3x01234567), voutput_zero_point);
208
209 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
210 vacc23x01234567 = _mm256_permute4x64_epi64(vacc23x01234567, _MM_SHUFFLE(3, 1, 2, 0));
211
212 __m256i vout = _mm256_packs_epi16(vacc01x01234567, vacc23x01234567);
213
214 vout = _mm256_max_epi8(vout, voutput_min);
215
216 __m128i vout_lo = _mm256_castsi256_si128(vout);
217 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
218
219 if (nc >= 8) {
220 _mm_storeh_pi((__m64*) c3, _mm_castsi128_ps(vout_hi));
221 _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
222 _mm_storel_epi64((__m128i*) c1, vout_hi);
223 _mm_storel_epi64((__m128i*) c0, vout_lo);
224
225 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
226 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
227 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
228 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
229
230 a = (const int8_t**restrict) ((uintptr_t) a - ks);
231
232 nc -= 8;
233 } else {
234 if (nc & 4) {
235 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout_hi, 2));
236 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout_lo, 2));
237 _mm_storeu_si32(c1, vout_hi);
238 _mm_storeu_si32(c0, vout_lo);
239
240 c3 += 4;
241 c2 += 4;
242 c1 += 4;
243 c0 += 4;
244
245 vout_lo = _mm_srli_epi64(vout_lo, 32);
246 vout_hi = _mm_srli_epi64(vout_hi, 32);
247 }
248 if (nc & 2) {
249 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout_hi, 4));
250 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout_lo, 4));
251 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout_hi, 0));
252 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
253
254 c3 += 2;
255 c2 += 2;
256 c1 += 2;
257 c0 += 2;
258
259 vout_lo = _mm_srli_epi32(vout_lo, 16);
260 vout_hi = _mm_srli_epi32(vout_hi, 16);
261 }
262 if (nc & 1) {
263 *c3 = (int8_t) _mm_extract_epi8(vout_hi, 8);
264 *c2 = (int8_t) _mm_extract_epi8(vout_lo, 8);
265 *c1 = (int8_t) _mm_extract_epi8(vout_hi, 0);
266 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
267 }
268
269 nc = 0;
270 }
271 } while (nc != 0);
272 }