Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-3x8c8-minmax-avx256skx.c

raw
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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_3x8c8__avx256skx(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58
59 const __m128i vmask = _mm_set1_epi8(0xF0);
60 XNN_FORCE_REALIZATION(vmask);
61 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
62 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
63 XNN_FORCE_REALIZATION(vmin);
64 XNN_FORCE_REALIZATION(vmax);
65
66 do {
67 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
68 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
69 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
70 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
71 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
72 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
73 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
74 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
75 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
76 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
77 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
78 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
79 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
80 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
81 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
82 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
83 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
84 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
85 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
86 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
87 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
88 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
89 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
90 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
91 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
92 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
93 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
94 w = (const int32_t*) w + 8;
95
96 size_t k = kc;
97 while (k >= 16 * sizeof(int8_t)) {
98 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
99 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
100 a0 += 8;
101 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
102 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
103 a1 += 8;
104 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
105 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
106 a2 += 8;
107
108 __m128i vb01 = _mm_load_si128((const __m128i*) w);
109 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
110 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
111 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
112
113 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
114 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
115 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
116 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
117 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
118 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
119 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
120
121 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
122 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
123 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
124 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
125 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
126 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
127 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
128
129 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
130 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
131 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
132 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
133 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
134 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
135 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
136
137 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
138 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
139 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
140
141 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
142 vxa0 = _mm256_cvtepi8_epi16(va0);
143 a0 += 8;
144 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
145 vxa1 = _mm256_cvtepi8_epi16(va1);
146 a1 += 8;
147 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
148 vxa2 = _mm256_cvtepi8_epi16(va2);
149 a2 += 8;
150
151 vbm01 = _mm_and_si128(vb01, vmask);
152 vxb01 = _mm256_cvtepi8_epi16(vbm01);
153
154 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
155 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
156 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
157 vbm23 = _mm_and_si128(vb23, vmask);
158 vxb23 = _mm256_cvtepi8_epi16(vbm23);
159
160 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
161 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
162 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
163 vbm45 = _mm_and_si128(vb45, vmask);
164 vxb45 = _mm256_cvtepi8_epi16(vbm45);
165
166 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
167 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
168 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
169 vbm67 = _mm_and_si128(vb67, vmask);
170 vxb67 = _mm256_cvtepi8_epi16(vbm67);
171
172 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
173 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
174 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
175
176 w = (const int8_t*) w + 64;
177 k -= 16 * sizeof(int8_t);
178 }
179
180 while (k >= 8 * sizeof(int8_t)) {
181 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
182 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
183 a0 += 8;
184 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
185 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
186 a1 += 8;
187 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
188 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
189 a2 += 8;
190
191 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
192 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
193 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
194 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
195
196 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
197 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
198 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
199 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
200 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
201 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
202 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
203
204 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
205 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
206 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
207 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
208 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
209 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
210 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
211
212 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
213 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
214 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
215 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
216 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
217 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
218 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
219
220 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
221 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
222 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
223
224 w = (const int8_t*) w + 64;
225 k -= 8 * sizeof(int8_t);
226 }
227
228 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
229 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
230 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
231 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
232 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
233 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
234
235 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
236 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
237 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
238
239 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
240 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
241 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
242 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
243
244 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
245 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
246 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
247 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
248 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
249 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
250 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
251 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
252 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
253
254 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
255 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
256 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
257
258 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
259 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
260 w = (const float*) w + 16;
261 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
262 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
263 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
264
265 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
266 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
267 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
268
269 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
270 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
271 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
272 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
273 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
274 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
275 if XNN_LIKELY(nc >= 8) {
276 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
277 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
278 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
279 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
280 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
281 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
282
283 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
284 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
285 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
286
287 nc -= 8;
288 } else {
289 // Prepare mask for valid 16-bit elements (depends on nc).
290 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
291 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
292 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
293 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
294 nc = 0;
295 }
296 } while (nc != 0);
297 }