Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x8c8-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_4x8c8__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 <= 4);
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
61 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
62 XNN_FORCE_REALIZATION(vsign_mask);
63 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
64 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
65 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
66 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
67 // XNN_FORCE_REALIZATION(vinput_zero_point);
68 // XNN_FORCE_REALIZATION(vinput_inv_scale);
69 // XNN_FORCE_REALIZATION(voutput_min);
70 // XNN_FORCE_REALIZATION(voutput_max);
71 do {
72 const __m256i vksum01234567 = _mm256_load_si256(w);
73 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
74 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
75 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
76 __m256i vacc1x0123 = vacc0x0123;
77 __m256i vacc1x4567 = vacc0x4567;
78 __m256i vacc2x0123 = vacc0x0123;
79 __m256i vacc2x4567 = vacc0x4567;
80 __m256i vacc3x0123 = vacc0x0123;
81 __m256i vacc3x4567 = vacc0x4567;
82 w = (const int32_t*) w + 8;
83
84 size_t p = ks;
85 do {
86 const int8_t* restrict a0 = a[0];
87 if XNN_UNPREDICTABLE(a0 != zero) {
88 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89 } else {
90 a0 = zero_data;
91 }
92 const int8_t* restrict a1 = a[1];
93 if XNN_UNPREDICTABLE(a1 != zero) {
94 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
95 } else {
96 a1 = zero_data;
97 }
98 const int8_t* restrict a2 = a[2];
99 if XNN_UNPREDICTABLE(a2 != zero) {
100 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
101 } else {
102 a2 = zero_data;
103 }
104 const int8_t* restrict a3 = a[3];
105 if XNN_UNPREDICTABLE(a3 != zero) {
106 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
107 } else {
108 a3 = zero_data;
109 }
110 a += 4;
111
112 size_t k = kc;
113 while (k >= 16 * sizeof(int8_t)) {
114 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
115 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
116 a0 += 16;
117 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
118 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
119 a1 += 16;
120 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
121 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
122 a2 += 16;
123 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
124 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
125 a3 += 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
132 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
133 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
134 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
135 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
136 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
137 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
138 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
139 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
140 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
141 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
142 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
143 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
144 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
145 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
146 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
147 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
148
149 w = (const int8_t*) w + 128;
150 k -= 16 * sizeof(int8_t);
151 }
152
153 if (k != 0) {
154 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
155 a0 += 8;
156 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
157 a1 += 8;
158 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
159 a2 += 8;
160 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
161 a3 += 8;
162
163 const __m256i vb01234567x0123 = _mm256_load_si256(w);
164 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
165
166 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
167 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
168 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
169 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
170 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
171 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
172 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
173 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
174
175 w = (const int8_t*) w + 64;
176 k -= 8 * sizeof(int8_t);
177 }
178
179 p -= 4 * sizeof(void*);
180 } while (p != 0);
181
182
183 // Add adjacent pairs
184 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
185 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
186 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
187 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
188 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
189 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
190 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
191 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
192
193 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
194 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
195 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
196 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
197
198 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
199 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
200 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
201 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
202
203 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
204 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
205 w = (const float*) w + 16;
206 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
207 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
208 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
209 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
210
211 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
212 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
213 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
214 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
215
216 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
217 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
218 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
219 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
220
221 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
222 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
223 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
224 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
225 if XNN_LIKELY(nc >= 8) {
226 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
227 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
228 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
229 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
230 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
231 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
232 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
233 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
234
235 a = (const int8_t**restrict) ((uintptr_t) a - ks);
236 nc -= 8;
237 } else {
238 if (nc & 4) {
239 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
240 c3 += 4;
241 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
242 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
243 c2 += 4;
244 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
245 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
246 c1 += 4;
247 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
248 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
249 c0 += 4;
250 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
251 }
252 if (nc & 2) {
253 _mm_storeu_si32(c3, vfp16out3x01234567);
254 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
255 c3 += 2;
256 _mm_storeu_si32(c2, vfp16out2x01234567);
257 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
258 c2 += 2;
259 _mm_storeu_si32(c1, vfp16out1x01234567);
260 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
261 c1 += 2;
262 _mm_storeu_si32(c0, vfp16out0x01234567);
263 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
264 c0 += 2;
265 }
266 if (nc & 1) {
267 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
268 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
269 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
270 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
271 }
272 nc = 0;
273 }
274 } while (nc != 0);
275 }