Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-6x8c4-minmax-avxvnni-u2-acc2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c4-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 #include <stdint.h>
12
13 #include <immintrin.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/gemm.h"
17 #include "xnnpack/intrinsics-polyfill.h"
18 #include "xnnpack/math.h"
19 #include "xnnpack/unaligned.h"
20 #include "xnnpack/prefetch.h"
21
22
23 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_6x8c4__avxvnni_u2_acc2(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 float* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_f32_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 <= 6);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 const int8_t* a0 = a;
47 float* c0 = c;
48 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
55 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 a2 = a1;
58 c2 = c1;
59 }
60 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
61 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 4) {
63 a3 = a2;
64 c3 = c2;
65 }
66 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
67 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 4) {
69 a4 = a3;
70 c4 = c3;
71 }
72 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
73 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
74 if XNN_UNPREDICTABLE(mr != 6) {
75 a5 = a4;
76 c5 = c4;
77 }
78
79 const __m256i vsign_mask =_mm256_set1_epi8(0x80);
80 XNN_FORCE_REALIZATION(vsign_mask);
81 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
82 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
83 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
84 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
85 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
86 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
87 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
88 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
89 // XNN_FORCE_REALIZATION(voutput_min);
90 // XNN_FORCE_REALIZATION(voutput_max);
91 do {
92 const __m256i vksum01234567 = _mm256_load_si256(w);
93 __m256i vacc0x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
94 __m256i vacc0x1x01234567 = _mm256_setzero_si256();
95 __m256i vacc1x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
96 __m256i vacc1x1x01234567 = _mm256_setzero_si256();
97 __m256i vacc2x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
98 __m256i vacc2x1x01234567 = _mm256_setzero_si256();
99 __m256i vacc3x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
100 __m256i vacc3x1x01234567 = _mm256_setzero_si256();
101 __m256i vacc4x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
102 __m256i vacc4x1x01234567 = _mm256_setzero_si256();
103 __m256i vacc5x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
104 __m256i vacc5x1x01234567 = _mm256_setzero_si256();
105 w = (const int32_t*) w + 8;
106
107 size_t k = kc;
108 while (k >= 8 * sizeof(int8_t)) {
109 __m256i va0x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 0));
110 __m256i va0x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 4));
111 a0 += 8;
112 __m256i va1x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 0));
113 __m256i va1x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 4));
114 a1 += 8;
115 __m256i va2x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 0));
116 __m256i va2x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 4));
117 a2 += 8;
118 __m256i va3x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 0));
119 __m256i va3x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 4));
120 a3 += 8;
121 __m256i va4x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 0));
122 __m256i va4x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 4));
123 a4 += 8;
124 __m256i va5x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5 + 0));
125 __m256i va5x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5 + 4));
126 a5 += 8;
127
128 va0x0x0123 = _mm256_xor_si256(va0x0x0123, vsign_mask);
129 va0x1x0123 = _mm256_xor_si256(va0x1x0123, vsign_mask);
130 va1x0x0123 = _mm256_xor_si256(va1x0x0123, vsign_mask);
131 va1x1x0123 = _mm256_xor_si256(va1x1x0123, vsign_mask);
132 va2x0x0123 = _mm256_xor_si256(va2x0x0123, vsign_mask);
133 va2x1x0123 = _mm256_xor_si256(va2x1x0123, vsign_mask);
134 va3x0x0123 = _mm256_xor_si256(va3x0x0123, vsign_mask);
135 va3x1x0123 = _mm256_xor_si256(va3x1x0123, vsign_mask);
136 va4x0x0123 = _mm256_xor_si256(va4x0x0123, vsign_mask);
137 va4x1x0123 = _mm256_xor_si256(va4x1x0123, vsign_mask);
138 va5x0x0123 = _mm256_xor_si256(va5x0x0123, vsign_mask);
139 va5x1x0123 = _mm256_xor_si256(va5x1x0123, vsign_mask);
140
141 const __m256i vb0x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
142 const __m256i vb1x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
143 w = (const int8_t*) w + 64;
144
145 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0x0123, vb0x01234567);
146 vacc0x1x01234567 = _mm256_dpbusd_avx_epi32(vacc0x1x01234567, va0x1x0123, vb1x01234567);
147 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0x0123, vb0x01234567);
148 vacc1x1x01234567 = _mm256_dpbusd_avx_epi32(vacc1x1x01234567, va1x1x0123, vb1x01234567);
149 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0x0123, vb0x01234567);
150 vacc2x1x01234567 = _mm256_dpbusd_avx_epi32(vacc2x1x01234567, va2x1x0123, vb1x01234567);
151 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0x0123, vb0x01234567);
152 vacc3x1x01234567 = _mm256_dpbusd_avx_epi32(vacc3x1x01234567, va3x1x0123, vb1x01234567);
153 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0x0123, vb0x01234567);
154 vacc4x1x01234567 = _mm256_dpbusd_avx_epi32(vacc4x1x01234567, va4x1x0123, vb1x01234567);
155 vacc5x0x01234567 = _mm256_dpbusd_avx_epi32(vacc5x0x01234567, va5x0x0123, vb0x01234567);
156 vacc5x1x01234567 = _mm256_dpbusd_avx_epi32(vacc5x1x01234567, va5x1x0123, vb1x01234567);
157
158 k -= 8 * sizeof(int8_t);
159 }
160 vacc0x0x01234567 = _mm256_add_epi32(vacc0x0x01234567, vacc0x1x01234567);
161 vacc1x0x01234567 = _mm256_add_epi32(vacc1x0x01234567, vacc1x1x01234567);
162 vacc2x0x01234567 = _mm256_add_epi32(vacc2x0x01234567, vacc2x1x01234567);
163 vacc3x0x01234567 = _mm256_add_epi32(vacc3x0x01234567, vacc3x1x01234567);
164 vacc4x0x01234567 = _mm256_add_epi32(vacc4x0x01234567, vacc4x1x01234567);
165 vacc5x0x01234567 = _mm256_add_epi32(vacc5x0x01234567, vacc5x1x01234567);
166 while (k != 0) {
167 __m256i va0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0));
168 a0 += 4;
169 __m256i va1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1));
170 a1 += 4;
171 __m256i va2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2));
172 a2 += 4;
173 __m256i va3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3));
174 a3 += 4;
175 __m256i va4x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4));
176 a4 += 4;
177 __m256i va5x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5));
178 a5 += 4;
179
180 va0x0123 = _mm256_xor_si256(va0x0123, vsign_mask);
181 va1x0123 = _mm256_xor_si256(va1x0123, vsign_mask);
182 va2x0123 = _mm256_xor_si256(va2x0123, vsign_mask);
183 va3x0123 = _mm256_xor_si256(va3x0123, vsign_mask);
184 va4x0123 = _mm256_xor_si256(va4x0123, vsign_mask);
185 va5x0123 = _mm256_xor_si256(va5x0123, vsign_mask);
186
187 const __m256i vb01234567 = _mm256_load_si256(w);
188
189 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0123, vb01234567);
190 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0123, vb01234567);
191 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0123, vb01234567);
192 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0123, vb01234567);
193 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0123, vb01234567);
194 vacc5x0x01234567 = _mm256_dpbusd_avx_epi32(vacc5x0x01234567, va5x0123, vb01234567);
195
196 w = (const int8_t*) w + 32;
197 k -= 4 * sizeof(int8_t);
198 }
199
200 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x0x01234567);
201 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x0x01234567);
202 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x0x01234567);
203 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x0x01234567);
204 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x0x01234567);
205 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x0x01234567);
206
207 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
208 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
209 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
210 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
211 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
212 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
213
214 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
215 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
216 w = (const float*) w + 16;
217
218 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
219 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
220 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
221 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
222 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
223 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
224
225 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
226 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
227 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
228 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
229 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
230 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
231
232 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
233 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
234 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
235 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
236 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
237 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
238
239 if XNN_LIKELY(nc >= 8) {
240 _mm256_storeu_ps(c0, vout0x01234567);
241 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
242 c0 = (float*) ((uintptr_t) c0 + cn_stride);
243 _mm256_storeu_ps(c1, vout1x01234567);
244 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
245 c1 = (float*) ((uintptr_t) c1 + cn_stride);
246 _mm256_storeu_ps(c2, vout2x01234567);
247 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
248 c2 = (float*) ((uintptr_t) c2 + cn_stride);
249 _mm256_storeu_ps(c3, vout3x01234567);
250 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
251 c3 = (float*) ((uintptr_t) c3 + cn_stride);
252 _mm256_storeu_ps(c4, vout4x01234567);
253 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
254 c4 = (float*) ((uintptr_t) c4 + cn_stride);
255 _mm256_storeu_ps(c5, vout5x01234567);
256 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
257 c5 = (float*) ((uintptr_t) c5 + cn_stride);
258 nc -= 8;
259 } else {
260 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
261 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
262 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
263 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
264 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
265 __m128 vout5x0123 = _mm256_castps256_ps128(vout5x01234567);
266 if (nc & 4) {
267 _mm_storeu_ps(c0, vout0x0123);
268 c0 += 4;
269 _mm_storeu_ps(c1, vout1x0123);
270 c1 += 4;
271 _mm_storeu_ps(c2, vout2x0123);
272 c2 += 4;
273 _mm_storeu_ps(c3, vout3x0123);
274 c3 += 4;
275 _mm_storeu_ps(c4, vout4x0123);
276 c4 += 4;
277 _mm_storeu_ps(c5, vout5x0123);
278 c5 += 4;
279 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
280 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
281 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
282 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
283 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
284 vout5x0123 = _mm256_extractf128_ps(vout5x01234567, 1);
285 }
286 if (nc & 2) {
287 _mm_storel_pi((__m64*) c0, vout0x0123);
288 c0 += 2;
289 _mm_storel_pi((__m64*) c1, vout1x0123);
290 c1 += 2;
291 _mm_storel_pi((__m64*) c2, vout2x0123);
292 c2 += 2;
293 _mm_storel_pi((__m64*) c3, vout3x0123);
294 c3 += 2;
295 _mm_storel_pi((__m64*) c4, vout4x0123);
296 c4 += 2;
297 _mm_storel_pi((__m64*) c5, vout5x0123);
298 c5 += 2;
299 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
300 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
301 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
302 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
303 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
304 vout5x0123 = _mm_movehl_ps(vout5x0123, vout5x0123);
305 }
306 if (nc & 1) {
307 _mm_store_ss(c0, vout0x0123);
308 _mm_store_ss(c1, vout1x0123);
309 _mm_store_ss(c2, vout2x0123);
310 _mm_store_ss(c3, vout3x0123);
311 _mm_store_ss(c4, vout4x0123);
312 _mm_store_ss(c5, vout5x0123);
313 }
314 nc = 0;
315 }
316 } while (nc != 0);
317 }