Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-avx256vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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_5x8c8__avx256vnni_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 <= 5);
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, 8 * 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 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63
64 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
65 XNN_FORCE_REALIZATION(vsign_mask);
66 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
67 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
68 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
69 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
70 // XNN_FORCE_REALIZATION(voutput_zero_point);
71 // XNN_FORCE_REALIZATION(voutput_min);
72 do {
73 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
74 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
75 __m256i vacc1x0123 = vacc0x0123;
76 __m256i vacc1x4567 = vacc0x4567;
77 __m256i vacc2x0123 = vacc0x0123;
78 __m256i vacc2x4567 = vacc0x4567;
79 __m256i vacc3x0123 = vacc0x0123;
80 __m256i vacc3x4567 = vacc0x4567;
81 __m256i vacc4x0123 = vacc0x0123;
82 __m256i vacc4x4567 = vacc0x4567;
83 w = (const int32_t*) w + 8;
84
85 size_t p = ks;
86 do {
87 const int8_t* restrict a0 = a[0];
88 if XNN_UNPREDICTABLE(a0 != zero) {
89 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
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 }
95 const int8_t* restrict a2 = a[2];
96 if XNN_UNPREDICTABLE(a2 != zero) {
97 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
98 }
99 const int8_t* restrict a3 = a[3];
100 if XNN_UNPREDICTABLE(a3 != zero) {
101 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
102 }
103 const int8_t* restrict a4 = a[4];
104 if XNN_UNPREDICTABLE(a4 != zero) {
105 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
106 }
107 a += 5;
108
109 size_t k = kc;
110 while (k >= 16 * sizeof(int8_t)) {
111 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
112 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
113 a0 += 16;
114 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
115 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
116 a1 += 16;
117 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
118 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
119 a2 += 16;
120 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
121 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
122 a3 += 16;
123 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
124 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
125 a4 += 16;
126
127 const __m256i vb01234567x0123 = _mm256_load_si256(w);
128 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
129 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
130 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
131 xnn_prefetch_to_l1((const int8_t*) w + 896);
132
133 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
134 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
135 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
136 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
137 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
138 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
139 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
140 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
141 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
142 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
143 xnn_prefetch_to_l1((const int8_t*) w + 960);
144 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
145 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
146 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
147 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
148 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
149 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
150 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
151 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
152 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
153 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
154
155 w = (const int8_t*) w + 128;
156 k -= 16 * sizeof(int8_t);
157 }
158
159 if (k != 0) {
160 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
161 a0 += 8;
162 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
163 a1 += 8;
164 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
165 a2 += 8;
166 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
167 a3 += 8;
168 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
169 a4 += 8;
170
171 const __m256i vb01234567x0123 = _mm256_load_si256(w);
172 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
173
174 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
175 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
176 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
177 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
178 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
179 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
180 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
181 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
182 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
183 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
184 xnn_prefetch_to_l1((const int8_t*) w + 960);
185
186 w = (const int8_t*) w + 64;
187 k -= 8 * sizeof(int8_t);
188 }
189
190 p -= 5 * sizeof(void*);
191 } while (p != 0);
192
193
194 // Add adjacent pairs
195 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
196 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
197 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
198 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
199 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
200 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
201 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
202 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
203 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
204 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
205
206 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
207 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
208 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
209 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
210 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
211
212 const __m256 vscale01234567 = _mm256_load_ps(w);
213 w = (const float*) w + 8;
214 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
215 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
216 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
217 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
218 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
219
220 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
221 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
222 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
223 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
224 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
225
226 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
227 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
228 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
229 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
230 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
231
232 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
233 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
234 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
235 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
236 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
237
238 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
239 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
240 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
241 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
242 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
243
244 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
245 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
246 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
247 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
248 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
249
250 if (nc >= 8) {
251 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
252 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
253 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
254 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
255 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
256 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
257 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
258 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
259 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
260 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
261 a = (const int8_t**restrict) ((uintptr_t) a - ks);
262 nc -= 8;
263 } else {
264 // Prepare mask for valid 8-bit elements (depends on nc).
265 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
266 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
267 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
268 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
269 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
270 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
271 nc = 0;
272 }
273 } while (nc != 0);
274 }