Debian Codegraph @

src/f16-f32acc-igemm/gen/f16-f32acc-igemm-4x16-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_4x16__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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint16_t) == 0);
37 assert(ks != 0);
38 assert(ks % (4 * 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
63 do {
64 __m256 vacc0x0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
65 __m256 vacc0x1 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
66 __m256 vacc1x0 = vacc0x0;
67 __m256 vacc1x1 = vacc0x1;
68 __m256 vacc2x0 = vacc0x0;
69 __m256 vacc2x1 = vacc0x1;
70 __m256 vacc3x0 = vacc0x0;
71 __m256 vacc3x1 = vacc0x1;
72 w = (const xnn_float16*) w + 16;
73
74 size_t p = ks;
75 do {
76 const uint16_t* restrict a0 = (const uint16_t*) a[0];
77 assert(a0 != NULL);
78 if XNN_UNPREDICTABLE(a0 != (const uint16_t*) zero) {
79 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const uint16_t* restrict a1 = (const uint16_t*) a[1];
82 assert(a1 != NULL);
83 if XNN_UNPREDICTABLE(a1 != (const uint16_t*) zero) {
84 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const uint16_t* restrict a2 = (const uint16_t*) a[2];
87 assert(a2 != NULL);
88 if XNN_UNPREDICTABLE(a2 != (const uint16_t*) zero) {
89 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
90 }
91 const uint16_t* restrict a3 = (const uint16_t*) a[3];
92 assert(a3 != NULL);
93 if XNN_UNPREDICTABLE(a3 != (const uint16_t*) zero) {
94 a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset);
95 }
96 a += 4;
97
98 size_t k = kc;
99 do {
100 const __m256 vb0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
101 const __m256 vb1 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
102 w = (const xnn_float16*) w + 16;
103
104 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
105 a0 += 1;
106 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
107 a1 += 1;
108 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
109 a2 += 1;
110 const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3));
111 a3 += 1;
112
113 vacc0x0 = _mm256_fmadd_ps(va0, vb0, vacc0x0);
114 vacc0x1 = _mm256_fmadd_ps(va0, vb1, vacc0x1);
115 vacc1x0 = _mm256_fmadd_ps(va1, vb0, vacc1x0);
116 vacc1x1 = _mm256_fmadd_ps(va1, vb1, vacc1x1);
117 vacc2x0 = _mm256_fmadd_ps(va2, vb0, vacc2x0);
118 vacc2x1 = _mm256_fmadd_ps(va2, vb1, vacc2x1);
119 vacc3x0 = _mm256_fmadd_ps(va3, vb0, vacc3x0);
120 vacc3x1 = _mm256_fmadd_ps(va3, vb1, vacc3x1);
121
122 k -= sizeof(uint16_t);
123 } while (k != 0);
124 p -= 4 * sizeof(void*);
125 } while (p != 0);
126
127 vacc0x0 = _mm256_max_ps(vacc0x0, vmin);
128 vacc1x0 = _mm256_max_ps(vacc1x0, vmin);
129 vacc2x0 = _mm256_max_ps(vacc2x0, vmin);
130 vacc3x0 = _mm256_max_ps(vacc3x0, vmin);
131 vacc0x1 = _mm256_max_ps(vacc0x1, vmin);
132 vacc1x1 = _mm256_max_ps(vacc1x1, vmin);
133 vacc2x1 = _mm256_max_ps(vacc2x1, vmin);
134 vacc3x1 = _mm256_max_ps(vacc3x1, vmin);
135
136 vacc0x0 = _mm256_min_ps(vacc0x0, vmax);
137 vacc1x0 = _mm256_min_ps(vacc1x0, vmax);
138 vacc2x0 = _mm256_min_ps(vacc2x0, vmax);
139 vacc3x0 = _mm256_min_ps(vacc3x0, vmax);
140 vacc0x1 = _mm256_min_ps(vacc0x1, vmax);
141 vacc1x1 = _mm256_min_ps(vacc1x1, vmax);
142 vacc2x1 = _mm256_min_ps(vacc2x1, vmax);
143 vacc3x1 = _mm256_min_ps(vacc3x1, vmax);
144
145 if XNN_LIKELY(nc >= 16) {
146 _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x0, _MM_FROUND_TO_NEAREST_INT));
147 _mm_storeu_si128((__m128i*) (c3 + 8), _mm256_cvtps_ph(vacc3x1, _MM_FROUND_TO_NEAREST_INT));
148 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
149 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT));
150 _mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x1, _MM_FROUND_TO_NEAREST_INT));
151 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
152 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT));
153 _mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x1, _MM_FROUND_TO_NEAREST_INT));
154 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
155 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT));
156 _mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x1, _MM_FROUND_TO_NEAREST_INT));
157 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
158
159 a = (const xnn_float16**restrict) ((uintptr_t) a - ks);
160 nc -= 16;
161 } else {
162 __m128i vh3x0 = _mm256_cvtps_ph(vacc3x0, _MM_FROUND_TO_NEAREST_INT);
163 __m128i vh2x0 = _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT);
164 __m128i vh1x0 = _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT);
165 __m128i vh0x0 = _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT);
166 if (nc & 8) {
167 _mm_storeu_si128((__m128i*) c3, vh3x0);
168 _mm_storeu_si128((__m128i*) c2, vh2x0);
169 _mm_storeu_si128((__m128i*) c1, vh1x0);
170 _mm_storeu_si128((__m128i*) c0, vh0x0);
171
172 vh3x0 = _mm256_cvtps_ph(vacc3x1, _MM_FROUND_TO_NEAREST_INT);
173 vh2x0 = _mm256_cvtps_ph(vacc2x1, _MM_FROUND_TO_NEAREST_INT);
174 vh1x0 = _mm256_cvtps_ph(vacc1x1, _MM_FROUND_TO_NEAREST_INT);
175 vh0x0 = _mm256_cvtps_ph(vacc0x1, _MM_FROUND_TO_NEAREST_INT);
176
177 c3 += 8;
178 c2 += 8;
179 c1 += 8;
180 c0 += 8;
181 }
182 if (nc & 4) {
183 _mm_storel_epi64((__m128i*) c3, vh3x0);
184 _mm_storel_epi64((__m128i*) c2, vh2x0);
185 _mm_storel_epi64((__m128i*) c1, vh1x0);
186 _mm_storel_epi64((__m128i*) c0, vh0x0);
187
188 vh3x0 = _mm_unpackhi_epi64(vh3x0, vh3x0);
189 vh2x0 = _mm_unpackhi_epi64(vh2x0, vh2x0);
190 vh1x0 = _mm_unpackhi_epi64(vh1x0, vh1x0);
191 vh0x0 = _mm_unpackhi_epi64(vh0x0, vh0x0);
192
193 c3 += 4;
194 c2 += 4;
195 c1 += 4;
196 c0 += 4;
197 }
198 if (nc & 2) {
199 _mm_storeu_si32(c3, vh3x0);
200 _mm_storeu_si32(c2, vh2x0);
201 _mm_storeu_si32(c1, vh1x0);
202 _mm_storeu_si32(c0, vh0x0);
203
204 vh3x0 = _mm_srli_epi64(vh3x0, 32);
205 vh2x0 = _mm_srli_epi64(vh2x0, 32);
206 vh1x0 = _mm_srli_epi64(vh1x0, 32);
207 vh0x0 = _mm_srli_epi64(vh0x0, 32);
208
209 c3 += 2;
210 c2 += 2;
211 c1 += 2;
212 c0 += 2;
213 }
214 if (nc & 1) {
215 *c3 = _mm_extract_epi16(vh3x0, 0);
216 *c2 = _mm_extract_epi16(vh2x0, 0);
217 *c1 = _mm_extract_epi16(vh1x0, 0);
218 *c0 = _mm_extract_epi16(vh0x0, 0);
219 }
220
221 nc = 0;
222 }
223 } while (nc != 0);
224 }