1 // Auto-generated file. Do not edit!
2 // Template: src/f16-igemm/avx2-broadcast.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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
17
xnn_f16_igemm_minmax_ukernel_5x16__avx2_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const void ** restrict a,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const void * zero,const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_5x16__avx2_broadcast(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const void**restrict a,
24 const void*restrict w,
25 void*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const void* zero,
30 const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 5);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint16_t) == 0);
37 assert(ks != 0);
38 assert(ks % (5 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(uint16_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 uint16_t* c0 = c;
45 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61
62 do {
63 __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
64 __m256 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
65 __m256 vacc1x01234567 = vacc0x01234567;
66 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
67 __m256 vacc2x01234567 = vacc0x01234567;
68 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
69 __m256 vacc3x01234567 = vacc0x01234567;
70 __m256 vacc3x89ABCDEF = vacc0x89ABCDEF;
71 __m256 vacc4x01234567 = vacc0x01234567;
72 __m256 vacc4x89ABCDEF = vacc0x89ABCDEF;
73 w = (const uint16_t*) w + 16;
74
75 size_t p = ks;
76 do {
77 const uint16_t* restrict a0 = (const uint16_t*) a[0];
78 assert(a0 != NULL);
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
81 }
82 const uint16_t* restrict a1 = (const uint16_t*) a[1];
83 assert(a1 != NULL);
84 if XNN_UNPREDICTABLE(a1 != zero) {
85 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
86 }
87 const uint16_t* restrict a2 = (const uint16_t*) a[2];
88 assert(a2 != NULL);
89 if XNN_UNPREDICTABLE(a2 != zero) {
90 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
91 }
92 const uint16_t* restrict a3 = (const uint16_t*) a[3];
93 assert(a3 != NULL);
94 if XNN_UNPREDICTABLE(a3 != zero) {
95 a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset);
96 }
97 const uint16_t* restrict a4 = (const uint16_t*) a[4];
98 assert(a4 != NULL);
99 if XNN_UNPREDICTABLE(a4 != zero) {
100 a4 = (const uint16_t*) ((uintptr_t) a4 + a_offset);
101 }
102 a += 5;
103
104 size_t k = kc;
105 do {
106 const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
107 const __m256 vb89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
108 w = (const uint16_t*) w + 16;
109
110 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
111 a0 += 1;
112 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
113 a1 += 1;
114 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
115 a2 += 1;
116 const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3));
117 a3 += 1;
118 const __m256 va4 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a4));
119 a4 += 1;
120
121 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
122 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF), _MM_FROUND_NO_EXC));
123 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
124 vacc1x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF), _MM_FROUND_NO_EXC));
125 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
126 vacc2x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF), _MM_FROUND_NO_EXC));
127 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb01234567, vacc3x01234567), _MM_FROUND_NO_EXC));
128 vacc3x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb89ABCDEF, vacc3x89ABCDEF), _MM_FROUND_NO_EXC));
129 vacc4x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va4, vb01234567, vacc4x01234567), _MM_FROUND_NO_EXC));
130 vacc4x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va4, vb89ABCDEF, vacc4x89ABCDEF), _MM_FROUND_NO_EXC));
131
132 k -= sizeof(uint16_t);
133 } while (k != 0);
134 p -= 5 * sizeof(void*);
135 } while (p != 0);
136
137 const __m256 vscale = _mm256_load_ps(params->avx.scale);
138 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc0x01234567, vscale), _MM_FROUND_NO_EXC));
139 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc1x01234567, vscale), _MM_FROUND_NO_EXC));
140 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc2x01234567, vscale), _MM_FROUND_NO_EXC));
141 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc3x01234567, vscale), _MM_FROUND_NO_EXC));
142 vacc4x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc4x01234567, vscale), _MM_FROUND_NO_EXC));
143 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc0x89ABCDEF, vscale), _MM_FROUND_NO_EXC));
144 vacc1x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc1x89ABCDEF, vscale), _MM_FROUND_NO_EXC));
145 vacc2x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc2x89ABCDEF, vscale), _MM_FROUND_NO_EXC));
146 vacc3x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc3x89ABCDEF, vscale), _MM_FROUND_NO_EXC));
147 vacc4x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc4x89ABCDEF, vscale), _MM_FROUND_NO_EXC));
148
149 const __m256 vmin = _mm256_load_ps(params->avx.min);
150 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
151 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
152 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
153 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
154 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
155 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
156 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
157 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
158 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
159 vacc4x89ABCDEF = _mm256_max_ps(vacc4x89ABCDEF, vmin);
160
161 const __m256 vmax = _mm256_load_ps(params->avx.max);
162 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
163 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
164 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
165 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
166 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
167 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
168 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
169 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
170 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
171 vacc4x89ABCDEF = _mm256_min_ps(vacc4x89ABCDEF, vmax);
172
173 if XNN_LIKELY(nc >= 16) {
174 _mm_storeu_si128((__m128i*) c4, _mm256_cvtps_ph(vacc4x01234567, _MM_FROUND_NO_EXC));
175 _mm_storeu_si128((__m128i*) (c4 + 8), _mm256_cvtps_ph(vacc4x89ABCDEF, _MM_FROUND_NO_EXC));
176 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
177 _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC));
178 _mm_storeu_si128((__m128i*) (c3 + 8), _mm256_cvtps_ph(vacc3x89ABCDEF, _MM_FROUND_NO_EXC));
179 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
180 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
181 _mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC));
182 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
183 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
184 _mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC));
185 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
186 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
187 _mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC));
188 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
189
190 a = (const void**restrict) ((uintptr_t) a - ks);
191 nc -= 16;
192 } else {
193 __m128i vh4x01234567 = _mm256_cvtps_ph(vacc4x01234567, _MM_FROUND_NO_EXC);
194 __m128i vh3x01234567 = _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC);
195 __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
196 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
197 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
198 if (nc & 8) {
199 _mm_storeu_si128((__m128i*) c4, vh4x01234567);
200 _mm_storeu_si128((__m128i*) c3, vh3x01234567);
201 _mm_storeu_si128((__m128i*) c2, vh2x01234567);
202 _mm_storeu_si128((__m128i*) c1, vh1x01234567);
203 _mm_storeu_si128((__m128i*) c0, vh0x01234567);
204
205 vh4x01234567 = _mm256_cvtps_ph(vacc4x89ABCDEF, _MM_FROUND_NO_EXC);
206 vh3x01234567 = _mm256_cvtps_ph(vacc3x89ABCDEF, _MM_FROUND_NO_EXC);
207 vh2x01234567 = _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC);
208 vh1x01234567 = _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC);
209 vh0x01234567 = _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC);
210
211 c4 += 8;
212 c3 += 8;
213 c2 += 8;
214 c1 += 8;
215 c0 += 8;
216 }
217 if (nc & 4) {
218 _mm_storel_epi64((__m128i*) c4, vh4x01234567);
219 _mm_storel_epi64((__m128i*) c3, vh3x01234567);
220 _mm_storel_epi64((__m128i*) c2, vh2x01234567);
221 _mm_storel_epi64((__m128i*) c1, vh1x01234567);
222 _mm_storel_epi64((__m128i*) c0, vh0x01234567);
223
224 vh4x01234567 = _mm_unpackhi_epi64(vh4x01234567, vh4x01234567);
225 vh3x01234567 = _mm_unpackhi_epi64(vh3x01234567, vh3x01234567);
226 vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
227 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
228 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
229
230 c4 += 4;
231 c3 += 4;
232 c2 += 4;
233 c1 += 4;
234 c0 += 4;
235 }
236 if (nc & 2) {
237 _mm_storeu_si32(c4, vh4x01234567);
238 _mm_storeu_si32(c3, vh3x01234567);
239 _mm_storeu_si32(c2, vh2x01234567);
240 _mm_storeu_si32(c1, vh1x01234567);
241 _mm_storeu_si32(c0, vh0x01234567);
242
243 vh4x01234567 = _mm_srli_epi64(vh4x01234567, 32);
244 vh3x01234567 = _mm_srli_epi64(vh3x01234567, 32);
245 vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
246 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
247 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
248
249 c4 += 2;
250 c3 += 2;
251 c2 += 2;
252 c1 += 2;
253 c0 += 2;
254 }
255 if (nc & 1) {
256 *c4 = _mm_extract_epi16(vh4x01234567, 0);
257 *c3 = _mm_extract_epi16(vh3x01234567, 0);
258 *c2 = _mm_extract_epi16(vh2x01234567, 0);
259 *c1 = _mm_extract_epi16(vh1x01234567, 0);
260 *c0 = _mm_extract_epi16(vh0x01234567, 0);
261 }
262
263 nc = 0;
264 }
265 } while (nc != 0);
266 }
267