Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x8c8-minmax-avxvnni-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_qd8_f16_qc8w_igemm_minmax_ukernel_4x8c8__avxvnni_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 xnn_float16* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 4);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 uint16_t* c0 = (uint16_t*) c;
49 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr != 4) {
59 c3 = c2;
60 }
61
62 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
63 XNN_FORCE_REALIZATION(vsign_mask);
64 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
65 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
66 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
67 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
68 // XNN_FORCE_REALIZATION(vinput_zero_point);
69 // XNN_FORCE_REALIZATION(vinput_inv_scale);
70 // XNN_FORCE_REALIZATION(voutput_min);
71 // XNN_FORCE_REALIZATION(voutput_max);
72 do {
73 const __m256i vksum01234567 = _mm256_load_si256(w);
74 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
75 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
76 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
77 __m256i vacc1x0123 = vacc0x0123;
78 __m256i vacc1x4567 = vacc0x4567;
79 __m256i vacc2x0123 = vacc0x0123;
80 __m256i vacc2x4567 = vacc0x4567;
81 __m256i vacc3x0123 = vacc0x0123;
82 __m256i vacc3x4567 = 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 } else {
91 a0 = zero_data;
92 }
93 const int8_t* restrict a1 = a[1];
94 if XNN_UNPREDICTABLE(a1 != zero) {
95 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
96 } else {
97 a1 = zero_data;
98 }
99 const int8_t* restrict a2 = a[2];
100 if XNN_UNPREDICTABLE(a2 != zero) {
101 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
102 } else {
103 a2 = zero_data;
104 }
105 const int8_t* restrict a3 = a[3];
106 if XNN_UNPREDICTABLE(a3 != zero) {
107 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
108 } else {
109 a3 = zero_data;
110 }
111 a += 4;
112
113 size_t k = kc;
114 while (k >= 16 * sizeof(int8_t)) {
115 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
116 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
117 a0 += 16;
118 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
119 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
120 a1 += 16;
121 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
122 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
123 a2 += 16;
124 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
125 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
126 a3 += 16;
127
128 const __m256i vb01234567x0123 = _mm256_load_si256(w);
129 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
130 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
131 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
132 xnn_prefetch_to_l1((const int8_t*) w + 896);
133
134 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
135 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
136 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
137 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
138 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
139 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
140 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
141 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
142 xnn_prefetch_to_l1((const int8_t*) w + 960);
143 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
144 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
145 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
146 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
147 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
148 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
149 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
150 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
151
152 w = (const int8_t*) w + 128;
153 k -= 16 * sizeof(int8_t);
154 }
155
156 if (k != 0) {
157 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
158 a0 += 8;
159 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
160 a1 += 8;
161 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
162 a2 += 8;
163 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
164 a3 += 8;
165
166 const __m256i vb01234567x0123 = _mm256_load_si256(w);
167 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
168
169 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
170 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
171 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
172 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
173 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
174 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
175 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
176 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
177 xnn_prefetch_to_l1((const int8_t*) w + 960);
178
179 w = (const int8_t*) w + 64;
180 k -= 8 * sizeof(int8_t);
181 }
182
183 p -= 4 * sizeof(void*);
184 } while (p != 0);
185
186
187 // Add adjacent pairs
188 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
189 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
190 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
191 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
192 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
193 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
194 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
195 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
196
197 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
198 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
199 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
200 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
201
202 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
203 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
204 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
205 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
206
207 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
208 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
209 w = (const float*) w + 16;
210 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
211 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
212 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
213 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
214
215 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
216 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
217 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
218 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
219
220 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
221 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
222 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
223 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
224
225 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
226 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
227 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
228 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
229 if XNN_LIKELY(nc >= 8) {
230 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
231 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
232 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
233 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
234 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
235 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
236 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
237 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
238
239 a = (const int8_t**restrict) ((uintptr_t) a - ks);
240 nc -= 8;
241 } else {
242 if (nc & 4) {
243 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
244 c3 += 4;
245 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
246 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
247 c2 += 4;
248 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
249 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
250 c1 += 4;
251 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
252 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
253 c0 += 4;
254 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
255 }
256 if (nc & 2) {
257 _mm_storeu_si32(c3, vfp16out3x01234567);
258 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
259 c3 += 2;
260 _mm_storeu_si32(c2, vfp16out2x01234567);
261 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
262 c2 += 2;
263 _mm_storeu_si32(c1, vfp16out1x01234567);
264 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
265 c1 += 2;
266 _mm_storeu_si32(c0, vfp16out0x01234567);
267 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
268 c0 += 2;
269 }
270 if (nc & 1) {
271 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
272 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
273 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
274 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
275 }
276 nc = 0;
277 }
278 } while (nc != 0);
279 }