Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-4x8c8-minmax-avx2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avx2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_gemm_minmax_ukernel_4x8c8__avx2(
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 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f16_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 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr != 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64
65 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
66 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
67 XNN_FORCE_REALIZATION(vmin);
68 XNN_FORCE_REALIZATION(vmax);
69
70 do {
71 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
72 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
73 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
74 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
75 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
76 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
77 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
78 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
79 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
80 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
81 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
82 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
83 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
84 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
85 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
86 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
87 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
88 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
89 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
90 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
91 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
92 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
93 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
94 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
95 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
96 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
97 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
98 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
99 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
100 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
101 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
102 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
103 w = (const int32_t*) w + 8;
104
105 size_t k = kc;
106
107 while (k >= 8 * sizeof(int8_t)) {
108 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
109 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
110 a0 += 8;
111 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
112 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
113 a1 += 8;
114 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
115 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
116 a2 += 8;
117 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
118 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
119 a3 += 8;
120
121 const __m256i vxb01 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) w));
122
123 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
124 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
125 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
126 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
127 const __m256i vxb23 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 16)));
128
129 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
130 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
131 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
132 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
133 const __m256i vxb45 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 32)));
134
135 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
136 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
137 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
138 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
139 const __m256i vxb67 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 48)));
140
141 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
142 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
143 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
144 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
145
146 w = (const int8_t*) w + 64;
147 k -= 8 * sizeof(int8_t);
148 }
149
150 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
151 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
152 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
153 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
154 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
155 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
156 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
157 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
158
159 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
160 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
161 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
162 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
163
164 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
165 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
166 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
167 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
168 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
169
170 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
171 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
172 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
173 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
174 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
175 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
176 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
177 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
178
179 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
180 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
181 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
182 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
183
184 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
185 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
186 w = (const float*) w + 16;
187 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
188 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
189 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
190 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
191
192 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
193 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
194 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
195 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
196
197 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
198 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
199 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
200 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
201 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
202 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
203 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
204 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
205 if XNN_LIKELY(nc >= 8) {
206 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
207 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
208 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
209 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
210 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
211 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
212 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
213 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
214
215 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
216 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
217 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
218 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
219
220 nc -= 8;
221 } else {
222 if (nc & 4) {
223 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
224 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
225 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
226 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
227
228 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
229 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
230 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
231 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
232
233 c0 += 4;
234 c1 += 4;
235 c2 += 4;
236 c3 += 4;
237 }
238 if (nc & 2) {
239 _mm_storeu_si32(c0, vfp16out0x01234567);
240 _mm_storeu_si32(c1, vfp16out1x01234567);
241 _mm_storeu_si32(c2, vfp16out2x01234567);
242 _mm_storeu_si32(c3, vfp16out3x01234567);
243
244 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
245 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
246 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
247 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
248
249 c0 += 2;
250 c1 += 2;
251 c2 += 2;
252 c3 += 2;
253 }
254 if (nc & 1) {
255 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
256 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
257 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
258 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
259 }
260 nc = 0;
261 }
262 } while (nc != 0);
263 }