Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x16c4-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_qs8_qc8w_igemm_minmax_fp32_ukernel_5x16c4__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 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 5);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62
63 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
64 XNN_FORCE_REALIZATION(vsign_mask);
65 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);
66 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
67 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
68 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
69 // XNN_FORCE_REALIZATION(voutput_zero_point);
70 // XNN_FORCE_REALIZATION(voutput_min);
71 do {
72 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
73 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
74 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
75 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
76 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
77 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
78 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
79 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
80 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
81 __m512i vacc1x4x0123456789ABCDEF = _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 }
90 const int8_t* restrict a1 = a[1];
91 if XNN_UNPREDICTABLE(a1 != zero) {
92 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
93 }
94 const int8_t* restrict a2 = a[2];
95 if XNN_UNPREDICTABLE(a2 != zero) {
96 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
97 }
98 const int8_t* restrict a3 = a[3];
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
101 }
102 const int8_t* restrict a4 = a[4];
103 if XNN_UNPREDICTABLE(a4 != zero) {
104 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
105 }
106 a += 5;
107
108 size_t k = kc;
109 while (k >= 8 * sizeof(int8_t)) {
110 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
111 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
112 a0 += 8;
113 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
114 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
115 a1 += 8;
116 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
117 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
118 a2 += 8;
119 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
120 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
121 a3 += 8;
122 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
123 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
124 a4 += 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 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
134 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
135 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
136 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
137 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
138 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
139
140 w = (const int8_t*) w + 128;
141 k -= 8 * sizeof(int8_t);
142 }
143
144 if (k != 0) {
145 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
146 a0 += 4;
147 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
148 a1 += 4;
149 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
150 a2 += 4;
151 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
152 a3 += 4;
153 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
154 a4 += 4;
155
156 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
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 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
163
164 w = (const int8_t*) w + 64;
165 k -= 4 * sizeof(int8_t);
166 }
167
168 p -= 5 * sizeof(void*);
169 } while (p != 0);
170
171 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
172 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
173 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
174 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
175 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
176
177 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
178 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
179 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
180 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
181 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
182
183 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
184 w = (const float*) w + 16;
185 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
186 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
187 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
188 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
189 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
190
191 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
192 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
193 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
194 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
195 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
196
197 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
198 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
199 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
200 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
201 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
202
203 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
204 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
205 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
206 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
207 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
208
209 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
210 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
211 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
212 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
213 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
214
215 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
216 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
217 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
218 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
219 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
220
221 if (nc >= 16) {
222 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
223 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
224 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
225 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
226 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
227 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
228 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
229 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
230 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
231 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
232 a = (const int8_t**restrict) ((uintptr_t) a - ks);
233 nc -= 16;
234 } else {
235 // Prepare mask for valid 8-bit elements (depends on nc).
236 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
237 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
238 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
239 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
240 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
241 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
242 nc = 0;
243 }
244 } while (nc != 0);
245 }