Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-5x16c8-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512vnni.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
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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_5x16c8__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 5);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % 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 float* c0 = c;
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 c3 = c2;
60 }
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65
66 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
67 XNN_FORCE_REALIZATION(vsign_mask);
68 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
69 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
70 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
71 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
72 // XNN_FORCE_REALIZATION(vinput_zero_point);
73 // XNN_FORCE_REALIZATION(vinput_inv_scale);
74 // XNN_FORCE_REALIZATION(voutput_min);
75 // XNN_FORCE_REALIZATION(voutput_max);
76 do {
77 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
78 const __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
79 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
80 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
81 __m512i vacc1x01234567 = vacc0x01234567;
82 __m512i vacc1x89ABCDEF = vacc0x89ABCDEF;
83 __m512i vacc2x01234567 = vacc0x01234567;
84 __m512i vacc2x89ABCDEF = vacc0x89ABCDEF;
85 __m512i vacc3x01234567 = vacc0x01234567;
86 __m512i vacc3x89ABCDEF = vacc0x89ABCDEF;
87 __m512i vacc4x01234567 = vacc0x01234567;
88 __m512i vacc4x89ABCDEF = vacc0x89ABCDEF;
89 w = (const int32_t*) w + 16;
90
91 size_t p = ks;
92 do {
93 const int8_t* restrict a0 = a[0];
94 if XNN_UNPREDICTABLE(a0 != zero) {
95 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
96 } else {
97 a0 = zero_data;
98 }
99 const int8_t* restrict a1 = a[1];
100 if XNN_UNPREDICTABLE(a1 != zero) {
101 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
102 } else {
103 a1 = zero_data;
104 }
105 const int8_t* restrict a2 = a[2];
106 if XNN_UNPREDICTABLE(a2 != zero) {
107 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
108 } else {
109 a2 = zero_data;
110 }
111 const int8_t* restrict a3 = a[3];
112 if XNN_UNPREDICTABLE(a3 != zero) {
113 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
114 } else {
115 a3 = zero_data;
116 }
117 const int8_t* restrict a4 = a[4];
118 if XNN_UNPREDICTABLE(a4 != zero) {
119 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
120 } else {
121 a4 = zero_data;
122 }
123 a += 5;
124
125 size_t k = kc;
126 while (k >= 16 * sizeof(int8_t)) {
127 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
128 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
129 a0 += 16;
130 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
131 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
132 a1 += 16;
133 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
134 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
135 a2 += 16;
136 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
137 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
138 a3 += 16;
139 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
140 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
141 a4 += 16;
142
143 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
144 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
145 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
146 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
147 xnn_prefetch_to_l1((const int8_t*) w + 768);
148 xnn_prefetch_to_l1((const int8_t*) w + 832);
149
150 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
151 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
152 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
153 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
154 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
155 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
156 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
157 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
158 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
159 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
160 xnn_prefetch_to_l1((const int8_t*) w + 896);
161 xnn_prefetch_to_l1((const int8_t*) w + 960);
162 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
163 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
164 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
165 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
166 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
167 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
168 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
169 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
170 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
171 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
172
173 w = (const int8_t*) w + 256;
174 k -= 16 * sizeof(int8_t);
175 }
176
177 if (k != 0) {
178 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
179 a0 += 8;
180 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
181 a1 += 8;
182 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
183 a2 += 8;
184 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
185 a3 += 8;
186 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
187 a4 += 8;
188
189 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
190 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
191
192 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
193 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
194 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
195 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
196 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
197 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
198 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
199 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
200 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
201 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
202 xnn_prefetch_to_l1((const int8_t*) w + 896);
203 xnn_prefetch_to_l1((const int8_t*) w + 960);
204
205 w = (const int8_t*) w + 128;
206 k -= 8 * sizeof(int8_t);
207 }
208
209 p -= 5 * sizeof(void*);
210 } while (p != 0);
211
212
213 // Add adjacent pairs
214 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
215 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
216 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
217 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
218 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
219 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
220 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
221 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
222 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
223 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
224 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
225 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
226 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
227 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
228 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
229 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
230
231 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
232 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
233 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
234 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
235 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
236
237 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
238 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
239 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
240 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
241 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
242
243 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
244 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
245 w = (const float*) w + 32;
246
247 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
248 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
249 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
250 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
251 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
252
253 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
254 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
255 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
256 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
257 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
258
259 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
260 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
261 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
262 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
263 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
264
265 if XNN_LIKELY(nc >= 16) {
266 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
267 c4 = (float*) ((uintptr_t) c4 + cn_stride);
268 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
269 c3 = (float*) ((uintptr_t) c3 + cn_stride);
270 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
271 c2 = (float*) ((uintptr_t) c2 + cn_stride);
272 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
273 c1 = (float*) ((uintptr_t) c1 + cn_stride);
274 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
275 c0 = (float*) ((uintptr_t) c0 + cn_stride);
276 a = (const int8_t**restrict) ((uintptr_t) a - ks);
277 nc -= 16;
278 } else {
279 // Prepare mask for valid 32-bit elements (depends on nc).
280 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
281 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
282 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
283 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
284 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
285 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
286 nc = 0;
287 }
288 } while (nc != 0);
289 }