Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-5x16c4-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_5x16c4__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, 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
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 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
79 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
80 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
81 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
82 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
83 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
84 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
85 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
86 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
87 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
88 w = (const int32_t*) w + 16;
89
90 size_t p = ks;
91 do {
92 const int8_t* restrict a0 = a[0];
93 if XNN_UNPREDICTABLE(a0 != zero) {
94 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
95 } else {
96 a0 = zero_data;
97 }
98 const int8_t* restrict a1 = a[1];
99 if XNN_UNPREDICTABLE(a1 != zero) {
100 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101 } else {
102 a1 = zero_data;
103 }
104 const int8_t* restrict a2 = a[2];
105 if XNN_UNPREDICTABLE(a2 != zero) {
106 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
107 } else {
108 a2 = zero_data;
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 } else {
114 a3 = zero_data;
115 }
116 const int8_t* restrict a4 = a[4];
117 if XNN_UNPREDICTABLE(a4 != zero) {
118 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
119 } else {
120 a4 = zero_data;
121 }
122 a += 5;
123
124 size_t k = kc;
125 while (k >= 8 * sizeof(int8_t)) {
126 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
127 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
128 a0 += 8;
129 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
130 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
131 a1 += 8;
132 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
133 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
134 a2 += 8;
135 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
136 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
137 a3 += 8;
138 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
139 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
140 a4 += 8;
141
142 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
143 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
144 xnn_prefetch_to_l1((const int8_t*) w + 896);
145
146 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
147 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
148 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
149 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
150 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
151 xnn_prefetch_to_l1((const int8_t*) w + 960);
152 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
153 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
154 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
155 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
156 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
157
158 w = (const int8_t*) w + 128;
159 k -= 8 * sizeof(int8_t);
160 }
161
162 if (k != 0) {
163 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
164 a0 += 4;
165 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
166 a1 += 4;
167 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
168 a2 += 4;
169 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
170 a3 += 4;
171 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
172 a4 += 4;
173
174 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
175 xnn_prefetch_to_l1((const int8_t*) w + 960);
176
177 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
178 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
179 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
180 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
181 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
182
183 w = (const int8_t*) w + 64;
184 k -= 4 * sizeof(int8_t);
185 }
186
187 p -= 5 * sizeof(void*);
188 } while (p != 0);
189
190 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
191 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
192 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
193 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
194 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
195
196 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
197 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
198 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
199 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
200 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
201
202 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
203 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
204 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
205 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
206 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
207
208 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
209 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
210 w = (const float*) w + 32;
211
212 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
213 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
214 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
215 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
216 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
217
218 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
219 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
220 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
221 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
222 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
223
224 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
225 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
226 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
227 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
228 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
229
230 if XNN_LIKELY(nc >= 16) {
231 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
232 c4 = (float*) ((uintptr_t) c4 + cn_stride);
233 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
234 c3 = (float*) ((uintptr_t) c3 + cn_stride);
235 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
236 c2 = (float*) ((uintptr_t) c2 + cn_stride);
237 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
238 c1 = (float*) ((uintptr_t) c1 + cn_stride);
239 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
240 c0 = (float*) ((uintptr_t) c0 + cn_stride);
241 a = (const int8_t**restrict) ((uintptr_t) a - ks);
242 nc -= 16;
243 } else {
244 // Prepare mask for valid 32-bit elements (depends on nc).
245 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
246 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
247 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
248 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
249 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
250 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
251 nc = 0;
252 }
253 } while (nc != 0);
254 }