Debian Codegraph @

src/f16-f32acc-igemm/gen/f16-f32acc-igemm-5x16-minmax-avx2-broadcast.c

raw
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
18 void xnn_f16_f32acc_igemm_minmax_ukernel_5x16__avx2_broadcast(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const xnn_float16** restrict a,
24 const xnn_float16* restrict w,
25 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const xnn_float16* zero,
30 const union xnn_f16_minmax_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 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
45 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
46 XNN_FORCE_REALIZATION(vmin);
47 XNN_FORCE_REALIZATION(vmax);
48
49 uint16_t* c0 = (uint16_t*) c;
50 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 c1 = c0;
53 }
54 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 c2 = c1;
57 }
58 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr < 4) {
60 c3 = c2;
61 }
62 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 c4 = c3;
65 }
66
67 do {
68 __m256 vacc0x0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
69 __m256 vacc0x1 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
70 __m256 vacc1x0 = vacc0x0;
71 __m256 vacc1x1 = vacc0x1;
72 __m256 vacc2x0 = vacc0x0;
73 __m256 vacc2x1 = vacc0x1;
74 __m256 vacc3x0 = vacc0x0;
75 __m256 vacc3x1 = vacc0x1;
76 __m256 vacc4x0 = vacc0x0;
77 __m256 vacc4x1 = vacc0x1;
78 w = (const xnn_float16*) w + 16;
79
80 size_t p = ks;
81 do {
82 const uint16_t* restrict a0 = (const uint16_t*) a[0];
83 assert(a0 != NULL);
84 if XNN_UNPREDICTABLE(a0 != (const uint16_t*) zero) {
85 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
86 }
87 const uint16_t* restrict a1 = (const uint16_t*) a[1];
88 assert(a1 != NULL);
89 if XNN_UNPREDICTABLE(a1 != (const uint16_t*) zero) {
90 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
91 }
92 const uint16_t* restrict a2 = (const uint16_t*) a[2];
93 assert(a2 != NULL);
94 if XNN_UNPREDICTABLE(a2 != (const uint16_t*) zero) {
95 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
96 }
97 const uint16_t* restrict a3 = (const uint16_t*) a[3];
98 assert(a3 != NULL);
99 if XNN_UNPREDICTABLE(a3 != (const uint16_t*) zero) {
100 a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset);
101 }
102 const uint16_t* restrict a4 = (const uint16_t*) a[4];
103 assert(a4 != NULL);
104 if XNN_UNPREDICTABLE(a4 != (const uint16_t*) zero) {
105 a4 = (const uint16_t*) ((uintptr_t) a4 + a_offset);
106 }
107 a += 5;
108
109 size_t k = kc;
110 do {
111 const __m256 vb0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
112 const __m256 vb1 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
113 w = (const xnn_float16*) w + 16;
114
115 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
116 a0 += 1;
117 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
118 a1 += 1;
119 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
120 a2 += 1;
121 const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3));
122 a3 += 1;
123 const __m256 va4 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a4));
124 a4 += 1;
125
126 vacc0x0 = _mm256_fmadd_ps(va0, vb0, vacc0x0);
127 vacc0x1 = _mm256_fmadd_ps(va0, vb1, vacc0x1);
128 vacc1x0 = _mm256_fmadd_ps(va1, vb0, vacc1x0);
129 vacc1x1 = _mm256_fmadd_ps(va1, vb1, vacc1x1);
130 vacc2x0 = _mm256_fmadd_ps(va2, vb0, vacc2x0);
131 vacc2x1 = _mm256_fmadd_ps(va2, vb1, vacc2x1);
132 vacc3x0 = _mm256_fmadd_ps(va3, vb0, vacc3x0);
133 vacc3x1 = _mm256_fmadd_ps(va3, vb1, vacc3x1);
134 vacc4x0 = _mm256_fmadd_ps(va4, vb0, vacc4x0);
135 vacc4x1 = _mm256_fmadd_ps(va4, vb1, vacc4x1);
136
137 k -= sizeof(uint16_t);
138 } while (k != 0);
139 p -= 5 * sizeof(void*);
140 } while (p != 0);
141
142 vacc0x0 = _mm256_max_ps(vacc0x0, vmin);
143 vacc1x0 = _mm256_max_ps(vacc1x0, vmin);
144 vacc2x0 = _mm256_max_ps(vacc2x0, vmin);
145 vacc3x0 = _mm256_max_ps(vacc3x0, vmin);
146 vacc4x0 = _mm256_max_ps(vacc4x0, vmin);
147 vacc0x1 = _mm256_max_ps(vacc0x1, vmin);
148 vacc1x1 = _mm256_max_ps(vacc1x1, vmin);
149 vacc2x1 = _mm256_max_ps(vacc2x1, vmin);
150 vacc3x1 = _mm256_max_ps(vacc3x1, vmin);
151 vacc4x1 = _mm256_max_ps(vacc4x1, vmin);
152
153 vacc0x0 = _mm256_min_ps(vacc0x0, vmax);
154 vacc1x0 = _mm256_min_ps(vacc1x0, vmax);
155 vacc2x0 = _mm256_min_ps(vacc2x0, vmax);
156 vacc3x0 = _mm256_min_ps(vacc3x0, vmax);
157 vacc4x0 = _mm256_min_ps(vacc4x0, vmax);
158 vacc0x1 = _mm256_min_ps(vacc0x1, vmax);
159 vacc1x1 = _mm256_min_ps(vacc1x1, vmax);
160 vacc2x1 = _mm256_min_ps(vacc2x1, vmax);
161 vacc3x1 = _mm256_min_ps(vacc3x1, vmax);
162 vacc4x1 = _mm256_min_ps(vacc4x1, vmax);
163
164 if XNN_LIKELY(nc >= 16) {
165 _mm_storeu_si128((__m128i*) c4, _mm256_cvtps_ph(vacc4x0, _MM_FROUND_TO_NEAREST_INT));
166 _mm_storeu_si128((__m128i*) (c4 + 8), _mm256_cvtps_ph(vacc4x1, _MM_FROUND_TO_NEAREST_INT));
167 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
168 _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x0, _MM_FROUND_TO_NEAREST_INT));
169 _mm_storeu_si128((__m128i*) (c3 + 8), _mm256_cvtps_ph(vacc3x1, _MM_FROUND_TO_NEAREST_INT));
170 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
171 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT));
172 _mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x1, _MM_FROUND_TO_NEAREST_INT));
173 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
174 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT));
175 _mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x1, _MM_FROUND_TO_NEAREST_INT));
176 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
177 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT));
178 _mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x1, _MM_FROUND_TO_NEAREST_INT));
179 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
180
181 a = (const xnn_float16**restrict) ((uintptr_t) a - ks);
182 nc -= 16;
183 } else {
184 __m128i vh4x0 = _mm256_cvtps_ph(vacc4x0, _MM_FROUND_TO_NEAREST_INT);
185 __m128i vh3x0 = _mm256_cvtps_ph(vacc3x0, _MM_FROUND_TO_NEAREST_INT);
186 __m128i vh2x0 = _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT);
187 __m128i vh1x0 = _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT);
188 __m128i vh0x0 = _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT);
189 if (nc & 8) {
190 _mm_storeu_si128((__m128i*) c4, vh4x0);
191 _mm_storeu_si128((__m128i*) c3, vh3x0);
192 _mm_storeu_si128((__m128i*) c2, vh2x0);
193 _mm_storeu_si128((__m128i*) c1, vh1x0);
194 _mm_storeu_si128((__m128i*) c0, vh0x0);
195
196 vh4x0 = _mm256_cvtps_ph(vacc4x1, _MM_FROUND_TO_NEAREST_INT);
197 vh3x0 = _mm256_cvtps_ph(vacc3x1, _MM_FROUND_TO_NEAREST_INT);
198 vh2x0 = _mm256_cvtps_ph(vacc2x1, _MM_FROUND_TO_NEAREST_INT);
199 vh1x0 = _mm256_cvtps_ph(vacc1x1, _MM_FROUND_TO_NEAREST_INT);
200 vh0x0 = _mm256_cvtps_ph(vacc0x1, _MM_FROUND_TO_NEAREST_INT);
201
202 c4 += 8;
203 c3 += 8;
204 c2 += 8;
205 c1 += 8;
206 c0 += 8;
207 }
208 if (nc & 4) {
209 _mm_storel_epi64((__m128i*) c4, vh4x0);
210 _mm_storel_epi64((__m128i*) c3, vh3x0);
211 _mm_storel_epi64((__m128i*) c2, vh2x0);
212 _mm_storel_epi64((__m128i*) c1, vh1x0);
213 _mm_storel_epi64((__m128i*) c0, vh0x0);
214
215 vh4x0 = _mm_unpackhi_epi64(vh4x0, vh4x0);
216 vh3x0 = _mm_unpackhi_epi64(vh3x0, vh3x0);
217 vh2x0 = _mm_unpackhi_epi64(vh2x0, vh2x0);
218 vh1x0 = _mm_unpackhi_epi64(vh1x0, vh1x0);
219 vh0x0 = _mm_unpackhi_epi64(vh0x0, vh0x0);
220
221 c4 += 4;
222 c3 += 4;
223 c2 += 4;
224 c1 += 4;
225 c0 += 4;
226 }
227 if (nc & 2) {
228 _mm_storeu_si32(c4, vh4x0);
229 _mm_storeu_si32(c3, vh3x0);
230 _mm_storeu_si32(c2, vh2x0);
231 _mm_storeu_si32(c1, vh1x0);
232 _mm_storeu_si32(c0, vh0x0);
233
234 vh4x0 = _mm_srli_epi64(vh4x0, 32);
235 vh3x0 = _mm_srli_epi64(vh3x0, 32);
236 vh2x0 = _mm_srli_epi64(vh2x0, 32);
237 vh1x0 = _mm_srli_epi64(vh1x0, 32);
238 vh0x0 = _mm_srli_epi64(vh0x0, 32);
239
240 c4 += 2;
241 c3 += 2;
242 c2 += 2;
243 c1 += 2;
244 c0 += 2;
245 }
246 if (nc & 1) {
247 *c4 = _mm_extract_epi16(vh4x0, 0);
248 *c3 = _mm_extract_epi16(vh3x0, 0);
249 *c2 = _mm_extract_epi16(vh2x0, 0);
250 *c1 = _mm_extract_epi16(vh1x0, 0);
251 *c0 = _mm_extract_epi16(vh0x0, 0);
252 }
253
254 nc = 0;
255 }
256 } while (nc != 0);
257 }