Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-5x8c8-minmax-avx256skx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_5x8c8__avx256skx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 5);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (5 * sizeof(void*)) == 0);
42 assert(a_offset % 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 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65
66 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point);
67 const __m256 vinput_scale = _mm256_broadcast_ss(&quantization_params->inv_scale);
68 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
69 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
70 // XNN_FORCE_REALIZATION(vinput_zero_point);
71 // XNN_FORCE_REALIZATION(vinput_scale);
72 // XNN_FORCE_REALIZATION(vmin);
73 // XNN_FORCE_REALIZATION(vmax);
74
75 do {
76 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
77 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
78 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
79 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
80 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
81 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
82 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
83 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
84 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
85 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
86 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
87 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
88 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point);
89 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point);
90 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point);
91 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point);
92 __m256i vacc1x01 = vacc0x01;
93 __m256i vacc1x23 = vacc0x23;
94 __m256i vacc1x45 = vacc0x45;
95 __m256i vacc1x67 = vacc0x67;
96 __m256i vacc2x01 = vacc0x01;
97 __m256i vacc2x23 = vacc0x23;
98 __m256i vacc2x45 = vacc0x45;
99 __m256i vacc2x67 = vacc0x67;
100 __m256i vacc3x01 = vacc0x01;
101 __m256i vacc3x23 = vacc0x23;
102 __m256i vacc3x45 = vacc0x45;
103 __m256i vacc3x67 = vacc0x67;
104 __m256i vacc4x01 = vacc0x01;
105 __m256i vacc4x23 = vacc0x23;
106 __m256i vacc4x45 = vacc0x45;
107 __m256i vacc4x67 = vacc0x67;
108 w = (const int32_t*) w + 8;
109
110 size_t p = ks;
111 do {
112 const int8_t* restrict a0 = a[0];
113 if XNN_UNPREDICTABLE(a0 != zero) {
114 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
115 } else {
116 a0 = zero_data;
117 }
118 const int8_t* restrict a1 = a[1];
119 if XNN_UNPREDICTABLE(a1 != zero) {
120 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
121 } else {
122 a1 = zero_data;
123 }
124 const int8_t* restrict a2 = a[2];
125 if XNN_UNPREDICTABLE(a2 != zero) {
126 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
127 } else {
128 a2 = zero_data;
129 }
130 const int8_t* restrict a3 = a[3];
131 if XNN_UNPREDICTABLE(a3 != zero) {
132 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
133 } else {
134 a3 = zero_data;
135 }
136 const int8_t* restrict a4 = a[4];
137 if XNN_UNPREDICTABLE(a4 != zero) {
138 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
139 } else {
140 a4 = zero_data;
141 }
142 a += 5;
143
144 size_t k = 0;
145 while (k < kc) {
146 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
147 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
148 a0 += 8;
149 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
150 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
151 a1 += 8;
152 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
153 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
154 a2 += 8;
155 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
156 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
157 a3 += 8;
158 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
159 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
160 a4 += 8;
161
162 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
163 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
164
165 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
166 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
167 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
168 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
169 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
170 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
171 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
172
173 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
174 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
175 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
176 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
177 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
178 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
179 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
180
181 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
182 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
183 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
184 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
185 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
186 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
187 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
188
189 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
190 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
191 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
192 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
193 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
194
195 w = (const void*) ((const int8_t*) w + 64);
196 k += 8 * sizeof(int8_t);
197 }
198 p -= 5 * sizeof(void*);
199 } while (p != 0);
200
201 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
202 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
203 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
204 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
205 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
206 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
207 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
208 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
209 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
210 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
211
212 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
213 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
214 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
215 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
216 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
217
218 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
219 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
220 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
221 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
222 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
223 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
224
225 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
226 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
227 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
228 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
229 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
230
231 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale);
232 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale);
233 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale);
234 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale);
235 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale);
236
237 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
238 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
239 w = (const float*) w + 16;
240 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
241 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
242 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
243 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
244 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
245
246 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
247 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
248 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
249 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
250 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
251
252 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
253 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
254 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
255 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
256 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
257 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
258 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
259 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
260 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
261 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
262 if XNN_LIKELY(nc >= 8) {
263 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
264 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
265 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
266 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
267 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
268 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
269 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
270 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
271 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
272 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
273
274 a = (const int8_t**restrict) ((uintptr_t) a - ks);
275 nc -= 8;
276 } else {
277 // Prepare mask for valid 16-bit elements (depends on nc).
278 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
279 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
280 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
281 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
282 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
283 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
284 nc = 0;
285 }
286 } while (nc != 0);
287 }