Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-5x16c8-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-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_qb4w_gemm_minmax_ukernel_5x16c8__avx512vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 5);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 size_t bl = params->scalar.blocksize;
71 assert(bl != 0);
72 assert(bl <= kc);
73 assert(kc % bl == 0);
74 assert(bl % 32 == 0);
75
76 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
77 XNN_FORCE_REALIZATION(vsign_mask);
78 const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128);
79 const __m512 vinput_zero_point1 = _mm512_set1_ps((float) quantization_params[1].zero_point + 128);
80 const __m512 vinput_zero_point2 = _mm512_set1_ps((float) quantization_params[2].zero_point + 128);
81 const __m512 vinput_zero_point3 = _mm512_set1_ps((float) quantization_params[3].zero_point + 128);
82 const __m512 vinput_zero_point4 = _mm512_set1_ps((float) quantization_params[4].zero_point + 128);
83 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
84 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
85 // XNN_FORCE_REALIZATION(voutput_min);
86 // XNN_FORCE_REALIZATION(voutput_max);
87 const __m512i vmask = _mm512_set1_epi8(0xF0);
88 XNN_FORCE_REALIZATION(vmask);
89 do {
90 const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w);
91 __m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0);
92 __m512 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point1);
93 __m512 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point2);
94 __m512 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point3);
95 __m512 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point4);
96 w = (const int32_t*) w + 16;
97
98 for (size_t kb=0; kb < kc; kb+=bl) {
99 __m512i vacc0x01234567 = _mm512_setzero_epi32();
100 __m512i vacc0x89ABCDEF = _mm512_setzero_epi32();
101 __m512i vacc1x01234567 = _mm512_setzero_epi32();
102 __m512i vacc1x89ABCDEF = _mm512_setzero_epi32();
103 __m512i vacc2x01234567 = _mm512_setzero_epi32();
104 __m512i vacc2x89ABCDEF = _mm512_setzero_epi32();
105 __m512i vacc3x01234567 = _mm512_setzero_epi32();
106 __m512i vacc3x89ABCDEF = _mm512_setzero_epi32();
107 __m512i vacc4x01234567 = _mm512_setzero_epi32();
108 __m512i vacc4x89ABCDEF = _mm512_setzero_epi32();
109 size_t k = bl;
110 while (k >= 16 * sizeof(int8_t)) {
111 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
112 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
113 a0 += 16;
114 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
115 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
116 a1 += 16;
117 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
118 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
119 a2 += 16;
120 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
121 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
122 a3 += 16;
123 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
124 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
125 a4 += 16;
126
127 const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w);
128 const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64);
129 const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
130 const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
131 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask);
132 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
133 const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask);
134 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask);
135
136 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
137 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
138 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
139 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
140 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
141 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
142 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
143 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
144 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
145 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
146 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
147 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
148 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
149 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
150 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
151 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
152 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
153 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
154 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
155 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
156
157 w = (const int8_t*) w + 128;
158 k -= 16 * sizeof(int8_t);
159 }
160
161 if (k != 0) {
162 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
163 a0 += 8;
164 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
165 a1 += 8;
166 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
167 a2 += 8;
168 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
169 a3 += 8;
170 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
171 a4 += 8;
172
173 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
174 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
175 const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
176 const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
177
178 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
179 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
180 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
181 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
182 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
183 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
184 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
185 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
186 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
187 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
188
189 w = (const int8_t*) w + 128;
190 k -= 8 * sizeof(int8_t);
191 }
192 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32(
193 _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16));
194 w = (const uint16_t*) w + 16;
195
196 // Add adjacent pairs
197 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
198 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
199 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
200 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
201 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
202 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
203 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
204 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
205 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
206 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
207 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
208 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
209 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
210 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
211 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
212 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
213 __m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
214 __m512 vf1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
215 __m512 vf2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
216 __m512 vf3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
217 __m512 vf4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
218
219 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF);
220 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vf1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled1x0123456789ABCDEF);
221 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vf2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled2x0123456789ABCDEF);
222 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vf3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled3x0123456789ABCDEF);
223 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vf4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled4x0123456789ABCDEF);
224 }
225
226 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
227 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
228 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
229 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
230 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
231
232 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w);
233 w = (const float*) w + 16;
234
235 vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF);
236 vscaled1x0123456789ABCDEF = _mm512_add_ps(vscaled1x0123456789ABCDEF, vbias0123456789ABCDEF);
237 vscaled2x0123456789ABCDEF = _mm512_add_ps(vscaled2x0123456789ABCDEF, vbias0123456789ABCDEF);
238 vscaled3x0123456789ABCDEF = _mm512_add_ps(vscaled3x0123456789ABCDEF, vbias0123456789ABCDEF);
239 vscaled4x0123456789ABCDEF = _mm512_add_ps(vscaled4x0123456789ABCDEF, vbias0123456789ABCDEF);
240
241 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
242 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
243 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
244 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
245 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
246
247 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
248 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
249 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
250 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
251 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
252
253 if XNN_LIKELY(nc >= 16) {
254 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
255 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
256 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
257 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
258 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
259
260 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
261 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
262 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
263 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
264 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
265
266 c0 = (float*) ((uintptr_t) c0 + cn_stride);
267 c1 = (float*) ((uintptr_t) c1 + cn_stride);
268 c2 = (float*) ((uintptr_t) c2 + cn_stride);
269 c3 = (float*) ((uintptr_t) c3 + cn_stride);
270 c4 = (float*) ((uintptr_t) c4 + cn_stride);
271
272 nc -= 16;
273 } else {
274 // Prepare mask for valid 32-bit elements (depends on nc).
275 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
276 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
277 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
278 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
279 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
280 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
281 nc = 0;
282 }
283 } while (nc != 0);
284 }