Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x16c4-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_4x16c4__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 <= 4);
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
61 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
62 XNN_FORCE_REALIZATION(vsign_mask);
63 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
64 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
65 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
66 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
67 XNN_FORCE_REALIZATION(vinput_zero_point);
68 // XNN_FORCE_REALIZATION(vinput_inv_scale);
69 // XNN_FORCE_REALIZATION(voutput_min);
70 // XNN_FORCE_REALIZATION(voutput_max);
71 do {
72 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
73 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
74 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
75 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
76 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
77 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
78 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
79 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
80 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
81 w = (const int32_t*) w + 16;
82
83 size_t p = ks;
84 do {
85 const int8_t* restrict a0 = a[0];
86 if XNN_UNPREDICTABLE(a0 != zero) {
87 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
88 } else {
89 a0 = zero_data;
90 }
91 const int8_t* restrict a1 = a[1];
92 if XNN_UNPREDICTABLE(a1 != zero) {
93 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
94 } else {
95 a1 = zero_data;
96 }
97 const int8_t* restrict a2 = a[2];
98 if XNN_UNPREDICTABLE(a2 != zero) {
99 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
100 } else {
101 a2 = zero_data;
102 }
103 const int8_t* restrict a3 = a[3];
104 if XNN_UNPREDICTABLE(a3 != zero) {
105 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
106 } else {
107 a3 = zero_data;
108 }
109 a += 4;
110
111 size_t k = kc;
112 while (k >= 8 * sizeof(int8_t)) {
113 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
114 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
115 a0 += 8;
116 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
117 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
118 a1 += 8;
119 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
120 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
121 a2 += 8;
122 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
123 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
124 a3 += 8;
125
126 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
127 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
128
129 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
130 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
131 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
132 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
133 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
134 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
135 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
136 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
137
138 w = (const int8_t*) w + 128;
139 k -= 8 * sizeof(int8_t);
140 }
141
142 if (k != 0) {
143 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
144 a0 += 4;
145 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
146 a1 += 4;
147 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
148 a2 += 4;
149 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
150 a3 += 4;
151
152 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
153
154 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
155 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
156 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
157 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
158
159 w = (const int8_t*) w + 64;
160 k -= 4 * sizeof(int8_t);
161 }
162
163 p -= 4 * sizeof(void*);
164 } while (p != 0);
165
166 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
167 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
168 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
169 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
170
171 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
172 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
173 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
174 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
175
176 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
177 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
178 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
179 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
180
181 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
182 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
183 w = (const float*) w + 32;
184
185 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
186 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
187 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
188 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
189
190 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
191 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
192 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
193 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
194
195 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
196 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
197 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
198 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
199
200 if XNN_LIKELY(nc >= 16) {
201 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
202 c3 = (float*) ((uintptr_t) c3 + cn_stride);
203 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
204 c2 = (float*) ((uintptr_t) c2 + cn_stride);
205 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
206 c1 = (float*) ((uintptr_t) c1 + cn_stride);
207 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
208 c0 = (float*) ((uintptr_t) c0 + cn_stride);
209 a = (const int8_t**restrict) ((uintptr_t) a - ks);
210 nc -= 16;
211 } else {
212 // Prepare mask for valid 32-bit elements (depends on nc).
213 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
214 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
215 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
216 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
217 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
218 nc = 0;
219 }
220 } while (nc != 0);
221 }