Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 4);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
61 XNN_FORCE_REALIZATION(vsign_mask);
62 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
63 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
64 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
65 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
66 // XNN_FORCE_REALIZATION(voutput_zero_point);
67 // XNN_FORCE_REALIZATION(voutput_min);
68 do {
69 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
70 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
71 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
72 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
73 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
74 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
75 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
76 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
77 w = (const int32_t*) w + 16;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 }
85 const int8_t* restrict a1 = a[1];
86 if XNN_UNPREDICTABLE(a1 != zero) {
87 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
88 }
89 const int8_t* restrict a2 = a[2];
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
92 }
93 const int8_t* restrict a3 = a[3];
94 if XNN_UNPREDICTABLE(a3 != zero) {
95 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
96 }
97 a += 4;
98
99 size_t k = kc;
100 while (k >= 8 * sizeof(int8_t)) {
101 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
102 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
103 a0 += 8;
104 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
105 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
106 a1 += 8;
107 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
108 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
109 a2 += 8;
110 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
111 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
112 a3 += 8;
113
114 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
115 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
116 xnn_prefetch_to_l1((const int8_t*) w + 896);
117
118 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
119 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
120 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
121 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
122 xnn_prefetch_to_l1((const int8_t*) w + 960);
123 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
124 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
125 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
126 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
127
128 w = (const int8_t*) w + 128;
129 k -= 8 * sizeof(int8_t);
130 }
131
132 if (k != 0) {
133 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
134 a0 += 4;
135 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
136 a1 += 4;
137 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
138 a2 += 4;
139 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
140 a3 += 4;
141
142 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
143 xnn_prefetch_to_l1((const int8_t*) w + 960);
144
145 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
146 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
147 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
148 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
149
150 w = (const int8_t*) w + 64;
151 k -= 4 * sizeof(int8_t);
152 }
153
154 p -= 4 * sizeof(void*);
155 } while (p != 0);
156
157 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
158 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
159 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
160 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
161
162 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
163 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
164 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
165 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
166
167 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
168 w = (const float*) w + 16;
169 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
170 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
171 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
172 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
173
174 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
175 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
176 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
177 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
178
179 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
180 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
181 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
182 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
183
184 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
185 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
186 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
187 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
188
189 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
190 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
191 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
192 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
193
194 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
195 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
196 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
197 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
198
199 if (nc >= 16) {
200 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
201 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
202 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
203 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
204 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
205 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
206 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
207 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
208 a = (const int8_t**restrict) ((uintptr_t) a - ks);
209 nc -= 16;
210 } else {
211 // Prepare mask for valid 8-bit elements (depends on nc).
212 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
213 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
214 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
215 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
216 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
217 nc = 0;
218 }
219 } while (nc != 0);
220 }