Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-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_qd8_f32_qc8w_igemm_minmax_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 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 7);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 4 * sizeof(int8_t));
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72
73 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
74 XNN_FORCE_REALIZATION(vsign_mask);
75 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
76 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
77 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
78 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
79 XNN_FORCE_REALIZATION(vinput_zero_point);
80 // XNN_FORCE_REALIZATION(vinput_inv_scale);
81 // XNN_FORCE_REALIZATION(voutput_min);
82 // XNN_FORCE_REALIZATION(voutput_max);
83 do {
84 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
85 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
86 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
87 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
88 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
89 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
90 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
91 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
92 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
93 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
94 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
95 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
96 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
97 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
98 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
99 w = (const int32_t*) w + 16;
100
101 size_t p = ks;
102 do {
103 const int8_t* restrict a0 = a[0];
104 if XNN_UNPREDICTABLE(a0 != zero) {
105 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
106 } else {
107 a0 = zero_data;
108 }
109 const int8_t* restrict a1 = a[1];
110 if XNN_UNPREDICTABLE(a1 != zero) {
111 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
112 } else {
113 a1 = zero_data;
114 }
115 const int8_t* restrict a2 = a[2];
116 if XNN_UNPREDICTABLE(a2 != zero) {
117 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
118 } else {
119 a2 = zero_data;
120 }
121 const int8_t* restrict a3 = a[3];
122 if XNN_UNPREDICTABLE(a3 != zero) {
123 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
124 } else {
125 a3 = zero_data;
126 }
127 const int8_t* restrict a4 = a[4];
128 if XNN_UNPREDICTABLE(a4 != zero) {
129 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
130 } else {
131 a4 = zero_data;
132 }
133 const int8_t* restrict a5 = a[5];
134 if XNN_UNPREDICTABLE(a5 != zero) {
135 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
136 } else {
137 a5 = zero_data;
138 }
139 const int8_t* restrict a6 = a[6];
140 if XNN_UNPREDICTABLE(a6 != zero) {
141 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
142 } else {
143 a6 = zero_data;
144 }
145 a += 7;
146
147 size_t k = kc;
148 while (k >= 8 * sizeof(int8_t)) {
149 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
150 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
151 a0 += 8;
152 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
153 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
154 a1 += 8;
155 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
156 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
157 a2 += 8;
158 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
159 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
160 a3 += 8;
161 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
162 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
163 a4 += 8;
164 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
165 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
166 a5 += 8;
167 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
168 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
169 a6 += 8;
170
171 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
172 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
173
174 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
175 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
176 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
177 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
178 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
179 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
180 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
181 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
182 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
183 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
184 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
185 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
186 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
187 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
188
189 w = (const int8_t*) w + 128;
190 k -= 8 * sizeof(int8_t);
191 }
192
193 if (k != 0) {
194 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
195 a0 += 4;
196 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
197 a1 += 4;
198 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
199 a2 += 4;
200 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
201 a3 += 4;
202 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
203 a4 += 4;
204 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
205 a5 += 4;
206 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
207 a6 += 4;
208
209 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
210
211 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
212 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
213 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
214 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
215 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
216 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
217 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
218
219 w = (const int8_t*) w + 64;
220 k -= 4 * sizeof(int8_t);
221 }
222
223 p -= 7 * sizeof(void*);
224 } while (p != 0);
225
226 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
227 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
228 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
229 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
230 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
231 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
232 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
233
234 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
235 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
236 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
237 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
238 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
239 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
240 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
241
242 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
243 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
244 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
245 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
246 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
247 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
248 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
249
250 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
251 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
252 w = (const float*) w + 32;
253
254 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
255 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
256 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
257 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
258 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
259 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
260 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
261
262 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
263 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
264 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
265 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
266 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
267 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
268 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
269
270 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
271 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
272 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
273 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
274 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
275 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
276 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
277
278 if XNN_LIKELY(nc >= 16) {
279 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
280 c6 = (float*) ((uintptr_t) c6 + cn_stride);
281 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
282 c5 = (float*) ((uintptr_t) c5 + cn_stride);
283 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
284 c4 = (float*) ((uintptr_t) c4 + cn_stride);
285 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
286 c3 = (float*) ((uintptr_t) c3 + cn_stride);
287 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
288 c2 = (float*) ((uintptr_t) c2 + cn_stride);
289 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
290 c1 = (float*) ((uintptr_t) c1 + cn_stride);
291 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
292 c0 = (float*) ((uintptr_t) c0 + cn_stride);
293 a = (const int8_t**restrict) ((uintptr_t) a - ks);
294 nc -= 16;
295 } else {
296 // Prepare mask for valid 32-bit elements (depends on nc).
297 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
298 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
299 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
300 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
301 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
302 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
303 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
304 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
305 nc = 0;
306 }
307 } while (nc != 0);
308 }