Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x8c8-minmax-avx2.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_qd8_f16_qc8w_igemm_minmax_ukernel_4x8c8__avx2(
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 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 4);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (4 * sizeof(void*)) == 0);
42 assert(a_offset % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 uint16_t* c0 = (uint16_t*) c;
49 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr != 4) {
59 c3 = c2;
60 }
61
62 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point);
63 const __m256 vinput_scale = _mm256_broadcast_ss(&quantization_params->inv_scale);
64 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
65 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
66 // XNN_FORCE_REALIZATION(vinput_zero_point);
67 // XNN_FORCE_REALIZATION(vinput_scale);
68 // XNN_FORCE_REALIZATION(vmin);
69 // XNN_FORCE_REALIZATION(vmax);
70
71 do {
72 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
73 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
74 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
75 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
76 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
77 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
78 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
79 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
80 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
81 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
82 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
83 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
84 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point);
85 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point);
86 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point);
87 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point);
88 __m256i vacc1x01 = vacc0x01;
89 __m256i vacc1x23 = vacc0x23;
90 __m256i vacc1x45 = vacc0x45;
91 __m256i vacc1x67 = vacc0x67;
92 __m256i vacc2x01 = vacc0x01;
93 __m256i vacc2x23 = vacc0x23;
94 __m256i vacc2x45 = vacc0x45;
95 __m256i vacc2x67 = vacc0x67;
96 __m256i vacc3x01 = vacc0x01;
97 __m256i vacc3x23 = vacc0x23;
98 __m256i vacc3x45 = vacc0x45;
99 __m256i vacc3x67 = vacc0x67;
100 w = (const int32_t*) w + 8;
101
102 size_t p = ks;
103 do {
104 const int8_t* restrict a0 = a[0];
105 if XNN_UNPREDICTABLE(a0 != zero) {
106 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
107 } else {
108 a0 = zero_data;
109 }
110 const int8_t* restrict a1 = a[1];
111 if XNN_UNPREDICTABLE(a1 != zero) {
112 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
113 } else {
114 a1 = zero_data;
115 }
116 const int8_t* restrict a2 = a[2];
117 if XNN_UNPREDICTABLE(a2 != zero) {
118 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
119 } else {
120 a2 = zero_data;
121 }
122 const int8_t* restrict a3 = a[3];
123 if XNN_UNPREDICTABLE(a3 != zero) {
124 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
125 } else {
126 a3 = zero_data;
127 }
128 a += 4;
129
130 size_t k = 0;
131 while (k < kc) {
132 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
133 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
134 a0 += 8;
135 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
136 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
137 a1 += 8;
138 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
139 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
140 a2 += 8;
141 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
142 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
143 a3 += 8;
144
145 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
146 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
147
148 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
149 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
150 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
151 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
152 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
153 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
154
155 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
156 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
157 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
158 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
159 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
160 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
161
162 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
163 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
164 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
165 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
166 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
167 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
168
169 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
170 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
171 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
172 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
173
174 w = (const void*) ((const int8_t*) w + 64);
175 k += 8 * sizeof(int8_t);
176 }
177 p -= 4 * sizeof(void*);
178 } while (p != 0);
179
180 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
181 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
182 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
183 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
184 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
185 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
186 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
187 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
188
189 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
190 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
191 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
192 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
193
194 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
195 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
196 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
197 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
198 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
199
200 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
201 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
202 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
203 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
204
205 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale);
206 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale);
207 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale);
208 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale);
209
210 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
211 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
212 w = (const float*) w + 16;
213 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
214 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
215 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
216 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
217
218 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
219 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
220 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
221 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
222
223 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
224 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
225 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
226 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
227 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
228 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
229 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
230 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
231 if XNN_LIKELY(nc >= 8) {
232 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
233 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
234 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
235 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
236 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
237 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
238 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
239 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
240
241 a = (const int8_t**restrict) ((uintptr_t) a - ks);
242 nc -= 8;
243 } else {
244 if (nc & 4) {
245 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
246 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
247 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
248 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
249
250 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
251 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
252 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
253 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
254
255 c3 += 4;
256 c2 += 4;
257 c1 += 4;
258 c0 += 4;
259 }
260 if (nc & 2) {
261 _mm_storeu_si32(c3, vfp16out3x01234567);
262 _mm_storeu_si32(c2, vfp16out2x01234567);
263 _mm_storeu_si32(c1, vfp16out1x01234567);
264 _mm_storeu_si32(c0, vfp16out0x01234567);
265
266 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
267 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
268 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
269 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
270
271 c3 += 2;
272 c2 += 2;
273 c1 += 2;
274 c0 += 2;
275 }
276 if (nc & 1) {
277 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
278 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
279 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
280 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
281 }
282 nc = 0;
283 }
284 } while (nc != 0);
285 }