Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-7x8c8-minmax-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_qd8_f16_qc8w_igemm_minmax_ukernel_7x8c8__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 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 <= 7);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (7 * 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 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr < 6) {
67 c5 = c4;
68 }
69 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
70 if XNN_UNPREDICTABLE(mr <= 6) {
71 c6 = c5;
72 }
73
74 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point);
75 const __m256 vinput_scale = _mm256_broadcast_ss(&quantization_params->inv_scale);
76 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
77 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
78 // XNN_FORCE_REALIZATION(vinput_zero_point);
79 // XNN_FORCE_REALIZATION(vinput_scale);
80 // XNN_FORCE_REALIZATION(vmin);
81 // XNN_FORCE_REALIZATION(vmax);
82
83 do {
84 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
85 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
86 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
87 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
88 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
89 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
90 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
91 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
92 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
93 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
94 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
95 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
96 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point);
97 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point);
98 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point);
99 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point);
100 __m256i vacc1x01 = vacc0x01;
101 __m256i vacc1x23 = vacc0x23;
102 __m256i vacc1x45 = vacc0x45;
103 __m256i vacc1x67 = vacc0x67;
104 __m256i vacc2x01 = vacc0x01;
105 __m256i vacc2x23 = vacc0x23;
106 __m256i vacc2x45 = vacc0x45;
107 __m256i vacc2x67 = vacc0x67;
108 __m256i vacc3x01 = vacc0x01;
109 __m256i vacc3x23 = vacc0x23;
110 __m256i vacc3x45 = vacc0x45;
111 __m256i vacc3x67 = vacc0x67;
112 __m256i vacc4x01 = vacc0x01;
113 __m256i vacc4x23 = vacc0x23;
114 __m256i vacc4x45 = vacc0x45;
115 __m256i vacc4x67 = vacc0x67;
116 __m256i vacc5x01 = vacc0x01;
117 __m256i vacc5x23 = vacc0x23;
118 __m256i vacc5x45 = vacc0x45;
119 __m256i vacc5x67 = vacc0x67;
120 __m256i vacc6x01 = vacc0x01;
121 __m256i vacc6x23 = vacc0x23;
122 __m256i vacc6x45 = vacc0x45;
123 __m256i vacc6x67 = vacc0x67;
124 w = (const int32_t*) w + 8;
125
126 size_t p = ks;
127 do {
128 const int8_t* restrict a0 = a[0];
129 if XNN_UNPREDICTABLE(a0 != zero) {
130 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
131 } else {
132 a0 = zero_data;
133 }
134 const int8_t* restrict a1 = a[1];
135 if XNN_UNPREDICTABLE(a1 != zero) {
136 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
137 } else {
138 a1 = zero_data;
139 }
140 const int8_t* restrict a2 = a[2];
141 if XNN_UNPREDICTABLE(a2 != zero) {
142 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
143 } else {
144 a2 = zero_data;
145 }
146 const int8_t* restrict a3 = a[3];
147 if XNN_UNPREDICTABLE(a3 != zero) {
148 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
149 } else {
150 a3 = zero_data;
151 }
152 const int8_t* restrict a4 = a[4];
153 if XNN_UNPREDICTABLE(a4 != zero) {
154 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
155 } else {
156 a4 = zero_data;
157 }
158 const int8_t* restrict a5 = a[5];
159 if XNN_UNPREDICTABLE(a5 != zero) {
160 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
161 } else {
162 a5 = zero_data;
163 }
164 const int8_t* restrict a6 = a[6];
165 if XNN_UNPREDICTABLE(a6 != zero) {
166 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
167 } else {
168 a6 = zero_data;
169 }
170 a += 7;
171
172 size_t k = 0;
173 while (k < kc) {
174 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
175 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
176 a0 += 8;
177 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
178 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
179 a1 += 8;
180 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
181 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
182 a2 += 8;
183 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
184 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
185 a3 += 8;
186 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
187 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
188 a4 += 8;
189 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
190 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
191 a5 += 8;
192 const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
193 const __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
194 a6 += 8;
195
196 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
197 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
198
199 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
200 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
201 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
202 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
203 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
204 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
205 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
206 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
207 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
208
209 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
210 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
211 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
212 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
213 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
214 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
215 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
216 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
217 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
218
219 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
220 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
221 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
222 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
223 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
224 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
225 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
226 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
227 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
228
229 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
230 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
231 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
232 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
233 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
234 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
235 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
236
237 w = (const void*) ((const int8_t*) w + 64);
238 k += 8 * sizeof(int8_t);
239 }
240 p -= 7 * sizeof(void*);
241 } while (p != 0);
242
243 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
244 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
245 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
246 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
247 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
248 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
249 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
250 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
251 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
252 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
253 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
254 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
255 const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23);
256 const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67);
257
258 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
259 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
260 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
261 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
262 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
263 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
264 const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657);
265
266 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
267 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
268 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
269 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
270 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
271 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
272 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
273 __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask);
274
275 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
276 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
277 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
278 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
279 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
280 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
281 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
282
283 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale);
284 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale);
285 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale);
286 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale);
287 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale);
288 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale);
289 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale);
290
291 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
292 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
293 w = (const float*) w + 16;
294 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
295 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
296 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
297 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
298 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
299 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
300 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
301
302 vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin);
303 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
304 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
305 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
306 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
307 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
308 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
309
310 vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax);
311 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
312 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
313 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
314 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
315 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
316 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
317 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
318 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
319 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
320 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
321 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
322 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
323 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
324 if XNN_LIKELY(nc >= 8) {
325 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
326 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
327 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
328 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
329 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
330 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
331 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
332 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
333 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
334 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
335 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
336 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
337 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
338 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
339
340 a = (const int8_t**restrict) ((uintptr_t) a - ks);
341 nc -= 8;
342 } else {
343 // Prepare mask for valid 16-bit elements (depends on nc).
344 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
345 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
346 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
347 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
348 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
349 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
350 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
351 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
352 nc = 0;
353 }
354 } while (nc != 0);
355 }