Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-7x16c4-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-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_7x16c4__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 <= 7);
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, 4 * 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 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70
71 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
72 XNN_FORCE_REALIZATION(vsign_mask);
73 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);
74 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
75 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
76 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
77 // XNN_FORCE_REALIZATION(voutput_zero_point);
78 // XNN_FORCE_REALIZATION(voutput_min);
79 do {
80 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
81 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
82 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
83 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
84 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
85 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
86 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
87 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
88 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
89 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
90 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
91 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
92 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
93 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
94 w = (const int32_t*) w + 16;
95
96 size_t p = ks;
97 do {
98 const int8_t* restrict a0 = a[0];
99 if XNN_UNPREDICTABLE(a0 != zero) {
100 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
101 }
102 const int8_t* restrict a1 = a[1];
103 if XNN_UNPREDICTABLE(a1 != zero) {
104 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
105 }
106 const int8_t* restrict a2 = a[2];
107 if XNN_UNPREDICTABLE(a2 != zero) {
108 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
109 }
110 const int8_t* restrict a3 = a[3];
111 if XNN_UNPREDICTABLE(a3 != zero) {
112 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
113 }
114 const int8_t* restrict a4 = a[4];
115 if XNN_UNPREDICTABLE(a4 != zero) {
116 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
117 }
118 const int8_t* restrict a5 = a[5];
119 if XNN_UNPREDICTABLE(a5 != zero) {
120 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
121 }
122 const int8_t* restrict a6 = a[6];
123 if XNN_UNPREDICTABLE(a6 != zero) {
124 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
125 }
126 a += 7;
127
128 size_t k = kc;
129 while (k >= 8 * sizeof(int8_t)) {
130 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
131 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
132 a0 += 8;
133 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
134 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
135 a1 += 8;
136 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
137 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
138 a2 += 8;
139 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
140 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
141 a3 += 8;
142 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
143 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
144 a4 += 8;
145 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
146 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
147 a5 += 8;
148 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
149 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
150 a6 += 8;
151
152 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
153 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
154
155 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
156 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
157 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
158 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
159 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
160 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
161 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
162 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
163 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
164 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
165 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
166 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
167 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
168 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
169
170 w = (const int8_t*) w + 128;
171 k -= 8 * sizeof(int8_t);
172 }
173
174 if (k != 0) {
175 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
176 a0 += 4;
177 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
178 a1 += 4;
179 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
180 a2 += 4;
181 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
182 a3 += 4;
183 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
184 a4 += 4;
185 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
186 a5 += 4;
187 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
188 a6 += 4;
189
190 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
191
192 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
193 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
194 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
195 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
196 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
197 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
198 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
199
200 w = (const int8_t*) w + 64;
201 k -= 4 * sizeof(int8_t);
202 }
203
204 p -= 7 * sizeof(void*);
205 } while (p != 0);
206
207 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
208 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
209 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
210 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
211 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
212 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
213 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
214
215 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
216 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
217 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
218 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
219 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
220 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
221 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
222
223 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
224 w = (const float*) w + 16;
225 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
226 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
227 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
228 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
229 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
230 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
231 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
232
233 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
234 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
235 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
236 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
237 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
238 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
239 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
240
241 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
242 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
243 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
244 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
245 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
246 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
247 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
248
249 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
250 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
251 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
252 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
253 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
254 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
255 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
256
257 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
258 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
259 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
260 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
261 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
262 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
263 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
264
265 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
266 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
267 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
268 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
269 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
270 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
271 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
272
273 if (nc >= 16) {
274 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
275 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
276 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
277 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
278 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
279 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
280 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
281 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
282 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
283 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
284 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
285 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
286 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
287 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
288 a = (const int8_t**restrict) ((uintptr_t) a - ks);
289 nc -= 16;
290 } else {
291 // Prepare mask for valid 8-bit elements (depends on nc).
292 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
293 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
294 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
295 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
296 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
297 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
298 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
299 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
300 nc = 0;
301 }
302 } while (nc != 0);
303 }