Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-7x16c4-minmax-avx512vnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_7x16c4__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 <= 7);
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, 4 * 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 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr < 6) {
67 c5 = c4;
68 }
69 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
70 if XNN_UNPREDICTABLE(mr <= 6) {
71 c6 = c5;
72 }
73
74 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
75 XNN_FORCE_REALIZATION(vsign_mask);
76 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
77 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
78 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
79 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
80 XNN_FORCE_REALIZATION(vinput_zero_point);
81 // XNN_FORCE_REALIZATION(vinput_inv_scale);
82 // XNN_FORCE_REALIZATION(voutput_min);
83 // XNN_FORCE_REALIZATION(voutput_max);
84 do {
85 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
86 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
87 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
88 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
89 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
90 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
91 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
92 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
93 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
94 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
95 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
96 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
97 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
98 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
99 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
100 w = (const int32_t*) w + 16;
101
102 size_t p = ks;
103 do {
104 const int8_t* restrict a0 = a[0];
105 if XNN_UNPREDICTABLE(a0 != zero) {
106 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
107 } else {
108 a0 = zero_data;
109 }
110 const int8_t* restrict a1 = a[1];
111 if XNN_UNPREDICTABLE(a1 != zero) {
112 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
113 } else {
114 a1 = zero_data;
115 }
116 const int8_t* restrict a2 = a[2];
117 if XNN_UNPREDICTABLE(a2 != zero) {
118 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
119 } else {
120 a2 = zero_data;
121 }
122 const int8_t* restrict a3 = a[3];
123 if XNN_UNPREDICTABLE(a3 != zero) {
124 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
125 } else {
126 a3 = zero_data;
127 }
128 const int8_t* restrict a4 = a[4];
129 if XNN_UNPREDICTABLE(a4 != zero) {
130 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
131 } else {
132 a4 = zero_data;
133 }
134 const int8_t* restrict a5 = a[5];
135 if XNN_UNPREDICTABLE(a5 != zero) {
136 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
137 } else {
138 a5 = zero_data;
139 }
140 const int8_t* restrict a6 = a[6];
141 if XNN_UNPREDICTABLE(a6 != zero) {
142 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
143 } else {
144 a6 = zero_data;
145 }
146 a += 7;
147
148 size_t k = kc;
149 while (k >= 8 * sizeof(int8_t)) {
150 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
151 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
152 a0 += 8;
153 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
154 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
155 a1 += 8;
156 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
157 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
158 a2 += 8;
159 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
160 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
161 a3 += 8;
162 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
163 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
164 a4 += 8;
165 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
166 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
167 a5 += 8;
168 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
169 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
170 a6 += 8;
171
172 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
173 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
174 xnn_prefetch_to_l1((const int8_t*) w + 896);
175
176 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
177 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
178 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
179 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
180 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
181 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
182 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
183 xnn_prefetch_to_l1((const int8_t*) w + 960);
184 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
185 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
186 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
187 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
188 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
189 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
190 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
191
192 w = (const int8_t*) w + 128;
193 k -= 8 * sizeof(int8_t);
194 }
195
196 if (k != 0) {
197 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
198 a0 += 4;
199 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
200 a1 += 4;
201 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
202 a2 += 4;
203 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
204 a3 += 4;
205 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
206 a4 += 4;
207 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
208 a5 += 4;
209 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
210 a6 += 4;
211
212 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
213 xnn_prefetch_to_l1((const int8_t*) w + 960);
214
215 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
216 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
217 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
218 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
219 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
220 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
221 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
222
223 w = (const int8_t*) w + 64;
224 k -= 4 * sizeof(int8_t);
225 }
226
227 p -= 7 * sizeof(void*);
228 } while (p != 0);
229
230 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
231 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
232 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
233 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
234 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
235 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
236 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
237
238 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
239 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
240 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
241 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
242 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
243 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
244 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
245
246 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
247 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
248 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
249 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
250 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
251 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
252 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
253
254 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
255 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
256 w = (const float*) w + 32;
257
258 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
259 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
260 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
261 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
262 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
263 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
264 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
265
266 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
267 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
268 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
269 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
270 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
271 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
272 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
273
274 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
275 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
276 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
277 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
278 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
279 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
280 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
281
282 if XNN_LIKELY(nc >= 16) {
283 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
284 c6 = (float*) ((uintptr_t) c6 + cn_stride);
285 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
286 c5 = (float*) ((uintptr_t) c5 + cn_stride);
287 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
288 c4 = (float*) ((uintptr_t) c4 + cn_stride);
289 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
290 c3 = (float*) ((uintptr_t) c3 + cn_stride);
291 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
292 c2 = (float*) ((uintptr_t) c2 + cn_stride);
293 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
294 c1 = (float*) ((uintptr_t) c1 + cn_stride);
295 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
296 c0 = (float*) ((uintptr_t) c0 + cn_stride);
297 a = (const int8_t**restrict) ((uintptr_t) a - ks);
298 nc -= 16;
299 } else {
300 // Prepare mask for valid 32-bit elements (depends on nc).
301 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
302 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
303 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
304 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
305 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
306 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
307 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
308 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
309 nc = 0;
310 }
311 } while (nc != 0);
312 }