Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c8-minmax-avxvnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x8c8__avxvnni_prfm(
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 union xnn_f32_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 <= 4);
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
65 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
66 XNN_FORCE_REALIZATION(vsign_mask);
67 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
68 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
69 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
70 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
71 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
72 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
73 // XNN_FORCE_REALIZATION(voutput_min);
74 // XNN_FORCE_REALIZATION(voutput_max);
75 do {
76 const __m256i vksum01234567 = _mm256_load_si256(w);
77 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
78 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
79 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
80 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
81 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
82 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
83 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
84 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
85 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
86 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
87 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
88 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
89 w = (const int32_t*) w + 8;
90
91 size_t k = kc;
92 while (k >= 16 * sizeof(int8_t)) {
93 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
94 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
95 a0 += 16;
96 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
97 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
98 a1 += 16;
99 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
100 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
101 a2 += 16;
102 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
103 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
104 a3 += 16;
105
106 const __m256i vb01234567x0123 = _mm256_load_si256(w);
107 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
108 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
109 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
110 xnn_prefetch_to_l1((const int8_t*) w + 896);
111
112 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
113 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
114 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
115 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
116 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
117 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
118 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
119 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
120 xnn_prefetch_to_l1((const int8_t*) w + 960);
121 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
122 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
123 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
124 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
125 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
126 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
127 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
128 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
129
130 w = (const int8_t*) w + 128;
131 k -= 16 * sizeof(int8_t);
132 }
133
134 if (k != 0) {
135 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
136 a0 += 8;
137 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
138 a1 += 8;
139 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
140 a2 += 8;
141 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
142 a3 += 8;
143
144 const __m256i vb01234567x0123 = _mm256_load_si256(w);
145 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
146
147 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
148 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
149 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
150 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
151 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
152 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
153 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
154 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
155 xnn_prefetch_to_l1((const int8_t*) w + 960);
156
157 w = (const int8_t*) w + 64;
158 k -= 8 * sizeof(int8_t);
159 }
160
161 // Add adjacent pairs
162 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
163 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
164 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
165 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
166 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
167 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
168 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
169 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
170
171 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
172 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
173 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
174 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
175
176 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
177 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
178 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
179 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
180
181 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
182 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
183 w = (const float*) w + 16;
184
185 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
186 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
187 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
188 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
189
190 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
191 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
192 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
193 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
194
195 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
196 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
197 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
198 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
199
200 if XNN_LIKELY(nc >= 8) {
201 _mm256_storeu_ps(c0, vout0x01234567);
202 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
203 c0 = (float*) ((uintptr_t) c0 + cn_stride);
204 _mm256_storeu_ps(c1, vout1x01234567);
205 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
206 c1 = (float*) ((uintptr_t) c1 + cn_stride);
207 _mm256_storeu_ps(c2, vout2x01234567);
208 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
209 c2 = (float*) ((uintptr_t) c2 + cn_stride);
210 _mm256_storeu_ps(c3, vout3x01234567);
211 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
212 c3 = (float*) ((uintptr_t) c3 + cn_stride);
213 nc -= 8;
214 } else {
215 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
216 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
217 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
218 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
219 if (nc & 4) {
220 _mm_storeu_ps(c0, vout0x0123);
221 c0 += 4;
222 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
223 _mm_storeu_ps(c1, vout1x0123);
224 c1 += 4;
225 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
226 _mm_storeu_ps(c2, vout2x0123);
227 c2 += 4;
228 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
229 _mm_storeu_ps(c3, vout3x0123);
230 c3 += 4;
231 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
232 }
233 if (nc & 2) {
234 _mm_storel_pi((__m64*) c0, vout0x0123);
235 c0 += 2;
236 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
237 _mm_storel_pi((__m64*) c1, vout1x0123);
238 c1 += 2;
239 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
240 _mm_storel_pi((__m64*) c2, vout2x0123);
241 c2 += 2;
242 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
243 _mm_storel_pi((__m64*) c3, vout3x0123);
244 c3 += 2;
245 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
246 }
247 if (nc & 1) {
248 _mm_store_ss(c0, vout0x0123);
249 _mm_store_ss(c1, vout1x0123);
250 _mm_store_ss(c2, vout2x0123);
251 _mm_store_ss(c3, vout3x0123);
252 }
253 nc = 0;
254 }
255 } while (nc != 0);
256 }