Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-5x8c4-minmax-avxvnni-u2-acc2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c4-avxvnni.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 #include <stdint.h>
12
13 #include <immintrin.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/gemm.h"
17 #include "xnnpack/intrinsics-polyfill.h"
18 #include "xnnpack/math.h"
19 #include "xnnpack/unaligned.h"
20 #include "xnnpack/prefetch.h"
21
22
23 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_5x8c4__avxvnni_u2_acc2(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 float* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_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, 4 * sizeof(int8_t));
46 const int8_t* a0 = a;
47 float* c0 = c;
48 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
55 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 a2 = a1;
58 c2 = c1;
59 }
60 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
61 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 4) {
63 a3 = a2;
64 c3 = c2;
65 }
66 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
67 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 4) {
69 a4 = a3;
70 c4 = c3;
71 }
72
73 const __m256i vsign_mask =_mm256_set1_epi8(0x80);
74 XNN_FORCE_REALIZATION(vsign_mask);
75 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
76 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
77 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
78 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
79 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
80 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
81 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
82 // XNN_FORCE_REALIZATION(voutput_min);
83 // XNN_FORCE_REALIZATION(voutput_max);
84 do {
85 const __m256i vksum01234567 = _mm256_load_si256(w);
86 __m256i vacc0x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
87 __m256i vacc0x1x01234567 = _mm256_setzero_si256();
88 __m256i vacc1x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
89 __m256i vacc1x1x01234567 = _mm256_setzero_si256();
90 __m256i vacc2x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
91 __m256i vacc2x1x01234567 = _mm256_setzero_si256();
92 __m256i vacc3x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
93 __m256i vacc3x1x01234567 = _mm256_setzero_si256();
94 __m256i vacc4x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
95 __m256i vacc4x1x01234567 = _mm256_setzero_si256();
96 w = (const int32_t*) w + 8;
97
98 size_t k = kc;
99 while (k >= 8 * sizeof(int8_t)) {
100 __m256i va0x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 0));
101 __m256i va0x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 4));
102 a0 += 8;
103 __m256i va1x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 0));
104 __m256i va1x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 4));
105 a1 += 8;
106 __m256i va2x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 0));
107 __m256i va2x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 4));
108 a2 += 8;
109 __m256i va3x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 0));
110 __m256i va3x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 4));
111 a3 += 8;
112 __m256i va4x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 0));
113 __m256i va4x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 4));
114 a4 += 8;
115
116 va0x0x0123 = _mm256_xor_si256(va0x0x0123, vsign_mask);
117 va0x1x0123 = _mm256_xor_si256(va0x1x0123, vsign_mask);
118 va1x0x0123 = _mm256_xor_si256(va1x0x0123, vsign_mask);
119 va1x1x0123 = _mm256_xor_si256(va1x1x0123, vsign_mask);
120 va2x0x0123 = _mm256_xor_si256(va2x0x0123, vsign_mask);
121 va2x1x0123 = _mm256_xor_si256(va2x1x0123, vsign_mask);
122 va3x0x0123 = _mm256_xor_si256(va3x0x0123, vsign_mask);
123 va3x1x0123 = _mm256_xor_si256(va3x1x0123, vsign_mask);
124 va4x0x0123 = _mm256_xor_si256(va4x0x0123, vsign_mask);
125 va4x1x0123 = _mm256_xor_si256(va4x1x0123, vsign_mask);
126
127 const __m256i vb0x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
128 const __m256i vb1x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
129 w = (const int8_t*) w + 64;
130
131 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0x0123, vb0x01234567);
132 vacc0x1x01234567 = _mm256_dpbusd_avx_epi32(vacc0x1x01234567, va0x1x0123, vb1x01234567);
133 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0x0123, vb0x01234567);
134 vacc1x1x01234567 = _mm256_dpbusd_avx_epi32(vacc1x1x01234567, va1x1x0123, vb1x01234567);
135 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0x0123, vb0x01234567);
136 vacc2x1x01234567 = _mm256_dpbusd_avx_epi32(vacc2x1x01234567, va2x1x0123, vb1x01234567);
137 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0x0123, vb0x01234567);
138 vacc3x1x01234567 = _mm256_dpbusd_avx_epi32(vacc3x1x01234567, va3x1x0123, vb1x01234567);
139 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0x0123, vb0x01234567);
140 vacc4x1x01234567 = _mm256_dpbusd_avx_epi32(vacc4x1x01234567, va4x1x0123, vb1x01234567);
141
142 k -= 8 * sizeof(int8_t);
143 }
144 vacc0x0x01234567 = _mm256_add_epi32(vacc0x0x01234567, vacc0x1x01234567);
145 vacc1x0x01234567 = _mm256_add_epi32(vacc1x0x01234567, vacc1x1x01234567);
146 vacc2x0x01234567 = _mm256_add_epi32(vacc2x0x01234567, vacc2x1x01234567);
147 vacc3x0x01234567 = _mm256_add_epi32(vacc3x0x01234567, vacc3x1x01234567);
148 vacc4x0x01234567 = _mm256_add_epi32(vacc4x0x01234567, vacc4x1x01234567);
149 while (k != 0) {
150 __m256i va0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0));
151 a0 += 4;
152 __m256i va1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1));
153 a1 += 4;
154 __m256i va2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2));
155 a2 += 4;
156 __m256i va3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3));
157 a3 += 4;
158 __m256i va4x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4));
159 a4 += 4;
160
161 va0x0123 = _mm256_xor_si256(va0x0123, vsign_mask);
162 va1x0123 = _mm256_xor_si256(va1x0123, vsign_mask);
163 va2x0123 = _mm256_xor_si256(va2x0123, vsign_mask);
164 va3x0123 = _mm256_xor_si256(va3x0123, vsign_mask);
165 va4x0123 = _mm256_xor_si256(va4x0123, vsign_mask);
166
167 const __m256i vb01234567 = _mm256_load_si256(w);
168
169 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0123, vb01234567);
170 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0123, vb01234567);
171 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0123, vb01234567);
172 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0123, vb01234567);
173 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0123, vb01234567);
174
175 w = (const int8_t*) w + 32;
176 k -= 4 * sizeof(int8_t);
177 }
178
179 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x0x01234567);
180 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x0x01234567);
181 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x0x01234567);
182 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x0x01234567);
183 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x0x01234567);
184
185 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
186 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
187 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
188 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
189 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
190
191 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
192 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
193 w = (const float*) w + 16;
194
195 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
196 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
197 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
198 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
199 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
200
201 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
202 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
203 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
204 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
205 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
206
207 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
208 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
209 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
210 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
211 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
212
213 if XNN_LIKELY(nc >= 8) {
214 _mm256_storeu_ps(c0, vout0x01234567);
215 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
216 c0 = (float*) ((uintptr_t) c0 + cn_stride);
217 _mm256_storeu_ps(c1, vout1x01234567);
218 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
219 c1 = (float*) ((uintptr_t) c1 + cn_stride);
220 _mm256_storeu_ps(c2, vout2x01234567);
221 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
222 c2 = (float*) ((uintptr_t) c2 + cn_stride);
223 _mm256_storeu_ps(c3, vout3x01234567);
224 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
225 c3 = (float*) ((uintptr_t) c3 + cn_stride);
226 _mm256_storeu_ps(c4, vout4x01234567);
227 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
228 c4 = (float*) ((uintptr_t) c4 + cn_stride);
229 nc -= 8;
230 } else {
231 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
232 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
233 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
234 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
235 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
236 if (nc & 4) {
237 _mm_storeu_ps(c0, vout0x0123);
238 c0 += 4;
239 _mm_storeu_ps(c1, vout1x0123);
240 c1 += 4;
241 _mm_storeu_ps(c2, vout2x0123);
242 c2 += 4;
243 _mm_storeu_ps(c3, vout3x0123);
244 c3 += 4;
245 _mm_storeu_ps(c4, vout4x0123);
246 c4 += 4;
247 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
248 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
249 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
250 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
251 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
252 }
253 if (nc & 2) {
254 _mm_storel_pi((__m64*) c0, vout0x0123);
255 c0 += 2;
256 _mm_storel_pi((__m64*) c1, vout1x0123);
257 c1 += 2;
258 _mm_storel_pi((__m64*) c2, vout2x0123);
259 c2 += 2;
260 _mm_storel_pi((__m64*) c3, vout3x0123);
261 c3 += 2;
262 _mm_storel_pi((__m64*) c4, vout4x0123);
263 c4 += 2;
264 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
265 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
266 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
267 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
268 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
269 }
270 if (nc & 1) {
271 _mm_store_ss(c0, vout0x0123);
272 _mm_store_ss(c1, vout1x0123);
273 _mm_store_ss(c2, vout2x0123);
274 _mm_store_ss(c3, vout3x0123);
275 _mm_store_ss(c4, vout4x0123);
276 }
277 nc = 0;
278 }
279 } while (nc != 0);
280 }