Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x8c4-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_7x8c4__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 <= 7);
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 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
79 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
80 if XNN_UNPREDICTABLE(mr <= 6) {
81 a6 = a5;
82 c6 = c5;
83 }
84
85 const __m256i vsign_mask =_mm256_set1_epi8(0x80);
86 XNN_FORCE_REALIZATION(vsign_mask);
87 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
88 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
89 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
90 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
91 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
92 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
93 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
94 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
95 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
96 // XNN_FORCE_REALIZATION(voutput_min);
97 // XNN_FORCE_REALIZATION(voutput_max);
98 do {
99 const __m256i vksum01234567 = _mm256_load_si256(w);
100 __m256i vacc0x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
101 __m256i vacc0x1x01234567 = _mm256_setzero_si256();
102 __m256i vacc1x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
103 __m256i vacc1x1x01234567 = _mm256_setzero_si256();
104 __m256i vacc2x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
105 __m256i vacc2x1x01234567 = _mm256_setzero_si256();
106 __m256i vacc3x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
107 __m256i vacc3x1x01234567 = _mm256_setzero_si256();
108 __m256i vacc4x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
109 __m256i vacc4x1x01234567 = _mm256_setzero_si256();
110 __m256i vacc5x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
111 __m256i vacc5x1x01234567 = _mm256_setzero_si256();
112 __m256i vacc6x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
113 __m256i vacc6x1x01234567 = _mm256_setzero_si256();
114 w = (const int32_t*) w + 8;
115
116 size_t k = kc;
117 while (k >= 8 * sizeof(int8_t)) {
118 __m256i va0x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 0));
119 __m256i va0x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 4));
120 a0 += 8;
121 __m256i va1x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 0));
122 __m256i va1x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 4));
123 a1 += 8;
124 __m256i va2x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 0));
125 __m256i va2x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 4));
126 a2 += 8;
127 __m256i va3x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 0));
128 __m256i va3x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 4));
129 a3 += 8;
130 __m256i va4x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 0));
131 __m256i va4x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 4));
132 a4 += 8;
133 __m256i va5x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5 + 0));
134 __m256i va5x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5 + 4));
135 a5 += 8;
136 __m256i va6x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a6 + 0));
137 __m256i va6x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a6 + 4));
138 a6 += 8;
139
140 va0x0x0123 = _mm256_xor_si256(va0x0x0123, vsign_mask);
141 va0x1x0123 = _mm256_xor_si256(va0x1x0123, vsign_mask);
142 va1x0x0123 = _mm256_xor_si256(va1x0x0123, vsign_mask);
143 va1x1x0123 = _mm256_xor_si256(va1x1x0123, vsign_mask);
144 va2x0x0123 = _mm256_xor_si256(va2x0x0123, vsign_mask);
145 va2x1x0123 = _mm256_xor_si256(va2x1x0123, vsign_mask);
146 va3x0x0123 = _mm256_xor_si256(va3x0x0123, vsign_mask);
147 va3x1x0123 = _mm256_xor_si256(va3x1x0123, vsign_mask);
148 va4x0x0123 = _mm256_xor_si256(va4x0x0123, vsign_mask);
149 va4x1x0123 = _mm256_xor_si256(va4x1x0123, vsign_mask);
150 va5x0x0123 = _mm256_xor_si256(va5x0x0123, vsign_mask);
151 va5x1x0123 = _mm256_xor_si256(va5x1x0123, vsign_mask);
152 va6x0x0123 = _mm256_xor_si256(va6x0x0123, vsign_mask);
153 va6x1x0123 = _mm256_xor_si256(va6x1x0123, vsign_mask);
154
155 const __m256i vb0x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
156 const __m256i vb1x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
157 w = (const int8_t*) w + 64;
158
159 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0x0123, vb0x01234567);
160 vacc0x1x01234567 = _mm256_dpbusd_avx_epi32(vacc0x1x01234567, va0x1x0123, vb1x01234567);
161 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0x0123, vb0x01234567);
162 vacc1x1x01234567 = _mm256_dpbusd_avx_epi32(vacc1x1x01234567, va1x1x0123, vb1x01234567);
163 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0x0123, vb0x01234567);
164 vacc2x1x01234567 = _mm256_dpbusd_avx_epi32(vacc2x1x01234567, va2x1x0123, vb1x01234567);
165 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0x0123, vb0x01234567);
166 vacc3x1x01234567 = _mm256_dpbusd_avx_epi32(vacc3x1x01234567, va3x1x0123, vb1x01234567);
167 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0x0123, vb0x01234567);
168 vacc4x1x01234567 = _mm256_dpbusd_avx_epi32(vacc4x1x01234567, va4x1x0123, vb1x01234567);
169 vacc5x0x01234567 = _mm256_dpbusd_avx_epi32(vacc5x0x01234567, va5x0x0123, vb0x01234567);
170 vacc5x1x01234567 = _mm256_dpbusd_avx_epi32(vacc5x1x01234567, va5x1x0123, vb1x01234567);
171 vacc6x0x01234567 = _mm256_dpbusd_avx_epi32(vacc6x0x01234567, va6x0x0123, vb0x01234567);
172 vacc6x1x01234567 = _mm256_dpbusd_avx_epi32(vacc6x1x01234567, va6x1x0123, vb1x01234567);
173
174 k -= 8 * sizeof(int8_t);
175 }
176 vacc0x0x01234567 = _mm256_add_epi32(vacc0x0x01234567, vacc0x1x01234567);
177 vacc1x0x01234567 = _mm256_add_epi32(vacc1x0x01234567, vacc1x1x01234567);
178 vacc2x0x01234567 = _mm256_add_epi32(vacc2x0x01234567, vacc2x1x01234567);
179 vacc3x0x01234567 = _mm256_add_epi32(vacc3x0x01234567, vacc3x1x01234567);
180 vacc4x0x01234567 = _mm256_add_epi32(vacc4x0x01234567, vacc4x1x01234567);
181 vacc5x0x01234567 = _mm256_add_epi32(vacc5x0x01234567, vacc5x1x01234567);
182 vacc6x0x01234567 = _mm256_add_epi32(vacc6x0x01234567, vacc6x1x01234567);
183 while (k != 0) {
184 __m256i va0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0));
185 a0 += 4;
186 __m256i va1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1));
187 a1 += 4;
188 __m256i va2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2));
189 a2 += 4;
190 __m256i va3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3));
191 a3 += 4;
192 __m256i va4x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4));
193 a4 += 4;
194 __m256i va5x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5));
195 a5 += 4;
196 __m256i va6x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a6));
197 a6 += 4;
198
199 va0x0123 = _mm256_xor_si256(va0x0123, vsign_mask);
200 va1x0123 = _mm256_xor_si256(va1x0123, vsign_mask);
201 va2x0123 = _mm256_xor_si256(va2x0123, vsign_mask);
202 va3x0123 = _mm256_xor_si256(va3x0123, vsign_mask);
203 va4x0123 = _mm256_xor_si256(va4x0123, vsign_mask);
204 va5x0123 = _mm256_xor_si256(va5x0123, vsign_mask);
205 va6x0123 = _mm256_xor_si256(va6x0123, vsign_mask);
206
207 const __m256i vb01234567 = _mm256_load_si256(w);
208
209 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0123, vb01234567);
210 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0123, vb01234567);
211 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0123, vb01234567);
212 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0123, vb01234567);
213 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0123, vb01234567);
214 vacc5x0x01234567 = _mm256_dpbusd_avx_epi32(vacc5x0x01234567, va5x0123, vb01234567);
215 vacc6x0x01234567 = _mm256_dpbusd_avx_epi32(vacc6x0x01234567, va6x0123, vb01234567);
216
217 w = (const int8_t*) w + 32;
218 k -= 4 * sizeof(int8_t);
219 }
220
221 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x0x01234567);
222 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x0x01234567);
223 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x0x01234567);
224 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x0x01234567);
225 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x0x01234567);
226 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x0x01234567);
227 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x0x01234567);
228
229 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
230 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
231 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
232 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
233 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
234 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
235 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
236
237 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
238 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
239 w = (const float*) w + 16;
240
241 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
242 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
243 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
244 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
245 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
246 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
247 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
248
249 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
250 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
251 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
252 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
253 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
254 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
255 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
256
257 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
258 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
259 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
260 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
261 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
262 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
263 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
264
265 if XNN_LIKELY(nc >= 8) {
266 _mm256_storeu_ps(c0, vout0x01234567);
267 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
268 c0 = (float*) ((uintptr_t) c0 + cn_stride);
269 _mm256_storeu_ps(c1, vout1x01234567);
270 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
271 c1 = (float*) ((uintptr_t) c1 + cn_stride);
272 _mm256_storeu_ps(c2, vout2x01234567);
273 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
274 c2 = (float*) ((uintptr_t) c2 + cn_stride);
275 _mm256_storeu_ps(c3, vout3x01234567);
276 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
277 c3 = (float*) ((uintptr_t) c3 + cn_stride);
278 _mm256_storeu_ps(c4, vout4x01234567);
279 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
280 c4 = (float*) ((uintptr_t) c4 + cn_stride);
281 _mm256_storeu_ps(c5, vout5x01234567);
282 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
283 c5 = (float*) ((uintptr_t) c5 + cn_stride);
284 _mm256_storeu_ps(c6, vout6x01234567);
285 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
286 c6 = (float*) ((uintptr_t) c6 + cn_stride);
287 nc -= 8;
288 } else {
289 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
290 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
291 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
292 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
293 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
294 __m128 vout5x0123 = _mm256_castps256_ps128(vout5x01234567);
295 __m128 vout6x0123 = _mm256_castps256_ps128(vout6x01234567);
296 if (nc & 4) {
297 _mm_storeu_ps(c0, vout0x0123);
298 c0 += 4;
299 _mm_storeu_ps(c1, vout1x0123);
300 c1 += 4;
301 _mm_storeu_ps(c2, vout2x0123);
302 c2 += 4;
303 _mm_storeu_ps(c3, vout3x0123);
304 c3 += 4;
305 _mm_storeu_ps(c4, vout4x0123);
306 c4 += 4;
307 _mm_storeu_ps(c5, vout5x0123);
308 c5 += 4;
309 _mm_storeu_ps(c6, vout6x0123);
310 c6 += 4;
311 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
312 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
313 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
314 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
315 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
316 vout5x0123 = _mm256_extractf128_ps(vout5x01234567, 1);
317 vout6x0123 = _mm256_extractf128_ps(vout6x01234567, 1);
318 }
319 if (nc & 2) {
320 _mm_storel_pi((__m64*) c0, vout0x0123);
321 c0 += 2;
322 _mm_storel_pi((__m64*) c1, vout1x0123);
323 c1 += 2;
324 _mm_storel_pi((__m64*) c2, vout2x0123);
325 c2 += 2;
326 _mm_storel_pi((__m64*) c3, vout3x0123);
327 c3 += 2;
328 _mm_storel_pi((__m64*) c4, vout4x0123);
329 c4 += 2;
330 _mm_storel_pi((__m64*) c5, vout5x0123);
331 c5 += 2;
332 _mm_storel_pi((__m64*) c6, vout6x0123);
333 c6 += 2;
334 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
335 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
336 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
337 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
338 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
339 vout5x0123 = _mm_movehl_ps(vout5x0123, vout5x0123);
340 vout6x0123 = _mm_movehl_ps(vout6x0123, vout6x0123);
341 }
342 if (nc & 1) {
343 _mm_store_ss(c0, vout0x0123);
344 _mm_store_ss(c1, vout1x0123);
345 _mm_store_ss(c2, vout2x0123);
346 _mm_store_ss(c3, vout3x0123);
347 _mm_store_ss(c4, vout4x0123);
348 _mm_store_ss(c5, vout5x0123);
349 _mm_store_ss(c6, vout6x0123);
350 }
351 nc = 0;
352 }
353 } while (nc != 0);
354 }