Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x16c4-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_4x16c4__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 <= 4);
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
62 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
63 XNN_FORCE_REALIZATION(vsign_mask);
64 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
65 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
66 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
67 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
68 XNN_FORCE_REALIZATION(vinput_zero_point);
69 // XNN_FORCE_REALIZATION(vinput_inv_scale);
70 // XNN_FORCE_REALIZATION(voutput_min);
71 // XNN_FORCE_REALIZATION(voutput_max);
72 do {
73 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
74 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
75 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
76 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
77 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
78 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
79 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
80 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
81 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
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 } else {
90 a0 = zero_data;
91 }
92 const int8_t* restrict a1 = a[1];
93 if XNN_UNPREDICTABLE(a1 != zero) {
94 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
95 } else {
96 a1 = zero_data;
97 }
98 const int8_t* restrict a2 = a[2];
99 if XNN_UNPREDICTABLE(a2 != zero) {
100 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
101 } else {
102 a2 = zero_data;
103 }
104 const int8_t* restrict a3 = a[3];
105 if XNN_UNPREDICTABLE(a3 != zero) {
106 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
107 } else {
108 a3 = zero_data;
109 }
110 a += 4;
111
112 size_t k = kc;
113 while (k >= 8 * sizeof(int8_t)) {
114 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
115 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
116 a0 += 8;
117 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
118 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
119 a1 += 8;
120 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
121 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
122 a2 += 8;
123 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
124 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
125 a3 += 8;
126
127 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
128 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
129 xnn_prefetch_to_l1((const int8_t*) w + 896);
130
131 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
132 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
133 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
134 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
135 xnn_prefetch_to_l1((const int8_t*) w + 960);
136 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
137 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
138 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
139 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
140
141 w = (const int8_t*) w + 128;
142 k -= 8 * sizeof(int8_t);
143 }
144
145 if (k != 0) {
146 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
147 a0 += 4;
148 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
149 a1 += 4;
150 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
151 a2 += 4;
152 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
153 a3 += 4;
154
155 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
156 xnn_prefetch_to_l1((const int8_t*) w + 960);
157
158 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
159 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
160 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
161 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
162
163 w = (const int8_t*) w + 64;
164 k -= 4 * sizeof(int8_t);
165 }
166
167 p -= 4 * sizeof(void*);
168 } while (p != 0);
169
170 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
171 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
172 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
173 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
174
175 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
176 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
177 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
178 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
179
180 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
181 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
182 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
183 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
184
185 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
186 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
187 w = (const float*) w + 32;
188
189 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
190 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
191 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
192 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
193
194 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
195 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
196 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
197 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
198
199 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
200 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
201 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
202 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
203
204 if XNN_LIKELY(nc >= 16) {
205 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
206 c3 = (float*) ((uintptr_t) c3 + cn_stride);
207 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
208 c2 = (float*) ((uintptr_t) c2 + cn_stride);
209 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
210 c1 = (float*) ((uintptr_t) c1 + cn_stride);
211 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
212 c0 = (float*) ((uintptr_t) c0 + cn_stride);
213 a = (const int8_t**restrict) ((uintptr_t) a - ks);
214 nc -= 16;
215 } else {
216 // Prepare mask for valid 32-bit elements (depends on nc).
217 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
218 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
219 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
220 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
221 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
222 nc = 0;
223 }
224 } while (nc != 0);
225 }