Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x8c8-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_f32_qc8w_gemm_minmax_ukernel_7x8c8__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 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_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 <= 7);
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 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((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 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
71 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 6) {
73 a5 = a4;
74 c5 = c4;
75 }
76 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
77 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 6) {
79 a6 = a5;
80 c6 = c5;
81 }
82
83 const __m256 vmin = _mm256_set1_ps(params->scalar.min);
84 const __m256 vmax = _mm256_set1_ps(params->scalar.max);
85 XNN_FORCE_REALIZATION(vmin);
86 XNN_FORCE_REALIZATION(vmax);
87
88 do {
89 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
90 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
91 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
92 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
93 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
94 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
95 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
96 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
97 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
98 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
99 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
100 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
101 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
102 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
103 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
104 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
105 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
106 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
107 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
108 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
109 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
110 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
111 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
112 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
113 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
114 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
115 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
116 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
117 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
118 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
119 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
120 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
121 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
122 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
123 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
124 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
125 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
126 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point);
127 __m256i vacc5x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point5);
128 __m256i vacc5x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point5);
129 __m256i vacc5x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point5);
130 __m256i vacc5x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point5);
131 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point);
132 __m256i vacc6x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point6);
133 __m256i vacc6x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point6);
134 __m256i vacc6x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point6);
135 __m256i vacc6x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point6);
136 w = (const int32_t*) w + 8;
137
138 size_t k = kc;
139
140 while (k >= 8 * sizeof(int8_t)) {
141 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
142 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
143 a0 += 8;
144 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
145 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
146 a1 += 8;
147 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
148 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
149 a2 += 8;
150 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
151 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
152 a3 += 8;
153 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
154 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
155 a4 += 8;
156 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
157 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
158 a5 += 8;
159 const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
160 const __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
161 a6 += 8;
162
163 const __m256i vxb01 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) w));
164
165 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
166 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
167 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
168 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
169 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
170 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
171 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
172 const __m256i vxb23 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 16)));
173
174 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
175 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
176 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
177 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
178 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
179 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
180 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
181 const __m256i vxb45 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 32)));
182
183 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
184 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
185 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
186 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
187 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
188 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
189 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
190 const __m256i vxb67 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 48)));
191
192 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
193 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
194 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
195 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
196 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
197 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
198 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
199
200 w = (const int8_t*) w + 64;
201 k -= 8 * sizeof(int8_t);
202 }
203
204 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
205 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
206 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
207 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
208 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
209 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
210 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
211 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
212 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
213 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
214 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
215 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
216 const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23);
217 const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67);
218
219 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
220 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
221 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
222 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
223 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
224 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
225 const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657);
226
227 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
228 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
229 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
230 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
231 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
232 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
233 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
234 __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask);
235
236 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
237 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
238 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
239 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
240 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
241 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
242 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
243 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
244 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
245 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
246 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
247 const __m256 vinput_scale5 = _mm256_broadcast_ss(&quantization_params[5].inv_scale);
248 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
249 const __m256 vinput_scale6 = _mm256_broadcast_ss(&quantization_params[6].inv_scale);
250
251 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
252 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
253 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
254 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
255 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
256 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale5);
257 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale6);
258
259 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
260 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
261 w = (const float*) w + 16;
262 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
263 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
264 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
265 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
266 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
267 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
268 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
269
270 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
271 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
272 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
273 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
274 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
275 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
276 vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin);
277
278 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
279 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
280 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
281 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
282 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
283 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
284 vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax);
285
286 if XNN_LIKELY(nc >= 8) {
287 _mm256_storeu_ps(c0, vout0x01234567);
288 c0 = (float*) ((uintptr_t) c0 + cn_stride);
289 _mm256_storeu_ps(c1, vout1x01234567);
290 c1 = (float*) ((uintptr_t) c1 + cn_stride);
291 _mm256_storeu_ps(c2, vout2x01234567);
292 c2 = (float*) ((uintptr_t) c2 + cn_stride);
293 _mm256_storeu_ps(c3, vout3x01234567);
294 c3 = (float*) ((uintptr_t) c3 + cn_stride);
295 _mm256_storeu_ps(c4, vout4x01234567);
296 c4 = (float*) ((uintptr_t) c4 + cn_stride);
297 _mm256_storeu_ps(c5, vout5x01234567);
298 c5 = (float*) ((uintptr_t) c5 + cn_stride);
299 _mm256_storeu_ps(c6, vout6x01234567);
300 c6 = (float*) ((uintptr_t) c6 + cn_stride);
301
302 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
303 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
304 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
305 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
306 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
307 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
308 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
309
310 nc -= 8;
311 } else {
312 // Prepare mask for valid 32-bit elements (depends on nc).
313 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
314 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
315 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
316 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
317 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
318 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
319 _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
320 _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
321 nc = 0;
322 }
323 } while (nc != 0);
324 }