Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-5x8c8-minmax-avxvnni.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
20
21 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_5x8c8__avxvnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 5);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 uint16_t* c0 = (uint16_t*) c;
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64
65 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
66 XNN_FORCE_REALIZATION(vsign_mask);
67 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
68 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
69 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
70 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
71 // XNN_FORCE_REALIZATION(vinput_zero_point);
72 // XNN_FORCE_REALIZATION(vinput_inv_scale);
73 // XNN_FORCE_REALIZATION(voutput_min);
74 // XNN_FORCE_REALIZATION(voutput_max);
75 do {
76 const __m256i vksum01234567 = _mm256_load_si256(w);
77 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
78 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
79 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
80 __m256i vacc1x0123 = vacc0x0123;
81 __m256i vacc1x4567 = vacc0x4567;
82 __m256i vacc2x0123 = vacc0x0123;
83 __m256i vacc2x4567 = vacc0x4567;
84 __m256i vacc3x0123 = vacc0x0123;
85 __m256i vacc3x4567 = vacc0x4567;
86 __m256i vacc4x0123 = vacc0x0123;
87 __m256i vacc4x4567 = vacc0x4567;
88 w = (const int32_t*) w + 8;
89
90 size_t p = ks;
91 do {
92 const int8_t* restrict a0 = a[0];
93 if XNN_UNPREDICTABLE(a0 != zero) {
94 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
95 } else {
96 a0 = zero_data;
97 }
98 const int8_t* restrict a1 = a[1];
99 if XNN_UNPREDICTABLE(a1 != zero) {
100 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101 } else {
102 a1 = zero_data;
103 }
104 const int8_t* restrict a2 = a[2];
105 if XNN_UNPREDICTABLE(a2 != zero) {
106 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
107 } else {
108 a2 = zero_data;
109 }
110 const int8_t* restrict a3 = a[3];
111 if XNN_UNPREDICTABLE(a3 != zero) {
112 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
113 } else {
114 a3 = zero_data;
115 }
116 const int8_t* restrict a4 = a[4];
117 if XNN_UNPREDICTABLE(a4 != zero) {
118 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
119 } else {
120 a4 = zero_data;
121 }
122 a += 5;
123
124 size_t k = kc;
125 while (k >= 16 * sizeof(int8_t)) {
126 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
127 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
128 a0 += 16;
129 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
130 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
131 a1 += 16;
132 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
133 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
134 a2 += 16;
135 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
136 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
137 a3 += 16;
138 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
139 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
140 a4 += 16;
141
142 const __m256i vb01234567x0123 = _mm256_load_si256(w);
143 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
144 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
145 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
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 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
156 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
157 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
158 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
159 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
160 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
161 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
162 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
163 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
164 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
165 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
166 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
167
168 w = (const int8_t*) w + 128;
169 k -= 16 * sizeof(int8_t);
170 }
171
172 if (k != 0) {
173 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
174 a0 += 8;
175 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
176 a1 += 8;
177 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
178 a2 += 8;
179 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
180 a3 += 8;
181 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
182 a4 += 8;
183
184 const __m256i vb01234567x0123 = _mm256_load_si256(w);
185 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
186
187 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
188 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
189 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
190 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
191 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
192 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
193 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
194 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
195 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
196 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
197
198 w = (const int8_t*) w + 64;
199 k -= 8 * sizeof(int8_t);
200 }
201
202 p -= 5 * sizeof(void*);
203 } while (p != 0);
204
205
206 // Add adjacent pairs
207 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
208 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
209 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
210 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
211 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
212 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
213 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
214 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
215 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
216 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
217
218 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
219 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
220 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
221 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
222 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
223
224 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
225 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
226 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
227 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
228 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
229
230 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
231 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
232 w = (const float*) w + 16;
233 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
234 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
235 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
236 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
237 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
238
239 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
240 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
241 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
242 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
243 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
244
245 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
246 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
247 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
248 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
249 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
250
251 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
252 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
253 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
254 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
255 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
256 if XNN_LIKELY(nc >= 8) {
257 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
258 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
259 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
260 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
261 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
262 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
263 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
264 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
265 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
266 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
267
268 a = (const int8_t**restrict) ((uintptr_t) a - ks);
269 nc -= 8;
270 } else {
271 if (nc & 4) {
272 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
273 c4 += 4;
274 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
275 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
276 c3 += 4;
277 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
278 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
279 c2 += 4;
280 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
281 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
282 c1 += 4;
283 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
284 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
285 c0 += 4;
286 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
287 }
288 if (nc & 2) {
289 _mm_storeu_si32(c4, vfp16out4x01234567);
290 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
291 c4 += 2;
292 _mm_storeu_si32(c3, vfp16out3x01234567);
293 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
294 c3 += 2;
295 _mm_storeu_si32(c2, vfp16out2x01234567);
296 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
297 c2 += 2;
298 _mm_storeu_si32(c1, vfp16out1x01234567);
299 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
300 c1 += 2;
301 _mm_storeu_si32(c0, vfp16out0x01234567);
302 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
303 c0 += 2;
304 }
305 if (nc & 1) {
306 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
307 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
308 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
309 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
310 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
311 }
312 nc = 0;
313 }
314 } while (nc != 0);
315 }