Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x16c8-minmax-avx512vnni.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
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_5x16c8__avx512vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 5);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62
63 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
64 XNN_FORCE_REALIZATION(vsign_mask);
65 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
66 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
67 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
68 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
69 // XNN_FORCE_REALIZATION(voutput_zero_point);
70 XNN_FORCE_REALIZATION(voutput_min);
71 do {
72 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm256_load_si256((const __m256i*) w));
73 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm256_load_si256((const __m256i*) ((const int32_t*) w + 8)));
74 __m512i vacc1x01234567 = vacc0x01234567;
75 __m512i vacc1x89ABCDEF = vacc0x89ABCDEF;
76 __m512i vacc2x01234567 = vacc0x01234567;
77 __m512i vacc2x89ABCDEF = vacc0x89ABCDEF;
78 __m512i vacc3x01234567 = vacc0x01234567;
79 __m512i vacc3x89ABCDEF = vacc0x89ABCDEF;
80 __m512i vacc4x01234567 = vacc0x01234567;
81 __m512i vacc4x89ABCDEF = vacc0x89ABCDEF;
82 w = (const int32_t*) w + 16;
83
84 size_t p = ks;
85 do {
86 const int8_t* restrict a0 = a[0];
87 if XNN_UNPREDICTABLE(a0 != zero) {
88 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89 }
90 const int8_t* restrict a1 = a[1];
91 if XNN_UNPREDICTABLE(a1 != zero) {
92 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
93 }
94 const int8_t* restrict a2 = a[2];
95 if XNN_UNPREDICTABLE(a2 != zero) {
96 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
97 }
98 const int8_t* restrict a3 = a[3];
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
101 }
102 const int8_t* restrict a4 = a[4];
103 if XNN_UNPREDICTABLE(a4 != zero) {
104 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
105 }
106 a += 5;
107
108 size_t k = kc;
109 while (k >= 16 * sizeof(int8_t)) {
110 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
111 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
112 a0 += 16;
113 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
114 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
115 a1 += 16;
116 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
117 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
118 a2 += 16;
119 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
120 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
121 a3 += 16;
122 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
123 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
124 a4 += 16;
125
126 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
127 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
128 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
129 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
130
131 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
132 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
133 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
134 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
135 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
136 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
137 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
138 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
139 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
140 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
141 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
142 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
143 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
144 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
145 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
146 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
147 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
148 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
149 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
150 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
151
152 w = (const int8_t*) w + 256;
153 k -= 16 * sizeof(int8_t);
154 }
155
156 if (k != 0) {
157 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
158 a0 += 8;
159 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
160 a1 += 8;
161 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
162 a2 += 8;
163 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
164 a3 += 8;
165 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
166 a4 += 8;
167
168 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
169 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
170
171 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
172 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
173 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
174 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
175 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
176 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
177 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
178 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
179 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
180 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
181
182 w = (const int8_t*) w + 128;
183 k -= 8 * sizeof(int8_t);
184 }
185
186 p -= 5 * sizeof(void*);
187 } while (p != 0);
188
189
190 // Add adjacent pairs
191 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
192 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
193 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
194 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
195 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
196 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
197 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
198 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
199 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
200 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
201 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
202 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
203 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
204 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
205 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
206 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
207
208 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
209 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
210 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
211 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
212 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
213
214 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
215 w = (const float*) w + 16;
216 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
217 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
218 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
219 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
220 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
221
222 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
223 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
224 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
225 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
226 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
227
228 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
229 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
230 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
231 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
232 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
233
234 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
235 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
236 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
237 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
238 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
239
240 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
241 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
242 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
243 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
244 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
245
246 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
247 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
248 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
249 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
250 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
251
252 if (nc >= 16) {
253 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
254 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
255 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
256 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
257 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
258 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
259 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
260 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
261 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
262 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
263 a = (const int8_t**restrict) ((uintptr_t) a - ks);
264 nc -= 16;
265 } else {
266 // Prepare mask for valid 8-bit elements (depends on nc).
267 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
268 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
269 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
270 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
271 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
272 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
273 nc = 0;
274 }
275 } while (nc != 0);
276 }