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_4x16__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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_4x16__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_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 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
58 do {
59 __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
60 __m256 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
61 __m256 vacc1x01234567 = vacc0x01234567;
62 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
63 __m256 vacc2x01234567 = vacc0x01234567;
64 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
65 __m256 vacc3x01234567 = vacc0x01234567;
66 __m256 vacc3x89ABCDEF = vacc0x89ABCDEF;
67 w = (const uint16_t*) w + 16;
68
69 size_t p = ks;
70 do {
71 const uint16_t* restrict a0 = (const uint16_t*) a[0];
72 assert(a0 != NULL);
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
75 }
76 const uint16_t* restrict a1 = (const uint16_t*) a[1];
77 assert(a1 != NULL);
78 if XNN_UNPREDICTABLE(a1 != zero) {
79 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
80 }
81 const uint16_t* restrict a2 = (const uint16_t*) a[2];
82 assert(a2 != NULL);
83 if XNN_UNPREDICTABLE(a2 != zero) {
84 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
85 }
86 const uint16_t* restrict a3 = (const uint16_t*) a[3];
87 assert(a3 != NULL);
88 if XNN_UNPREDICTABLE(a3 != zero) {
89 a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset);
90 }
91 a += 4;
92
93 size_t k = kc;
94 do {
95 const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
96 const __m256 vb89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
97 w = (const uint16_t*) w + 16;
98
99 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
100 a0 += 1;
101 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
102 a1 += 1;
103 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
104 a2 += 1;
105 const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3));
106 a3 += 1;
107
108 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
109 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF), _MM_FROUND_NO_EXC));
110 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
111 vacc1x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF), _MM_FROUND_NO_EXC));
112 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
113 vacc2x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF), _MM_FROUND_NO_EXC));
114 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb01234567, vacc3x01234567), _MM_FROUND_NO_EXC));
115 vacc3x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb89ABCDEF, vacc3x89ABCDEF), _MM_FROUND_NO_EXC));
116
117 k -= sizeof(uint16_t);
118 } while (k != 0);
119 p -= 4 * sizeof(void*);
120 } while (p != 0);
121
122 const __m256 vmin = _mm256_load_ps(params->avx.min);
123 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
124 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
125 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
126 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
127 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
128 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
129 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
130 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
131
132 const __m256 vmax = _mm256_load_ps(params->avx.max);
133 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
134 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
135 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
136 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
137 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
138 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
139 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
140 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
141
142 if XNN_LIKELY(nc >= 16) {
143 _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC));
144 _mm_storeu_si128((__m128i*) (c3 + 8), _mm256_cvtps_ph(vacc3x89ABCDEF, _MM_FROUND_NO_EXC));
145 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
146 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
147 _mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC));
148 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
149 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
150 _mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC));
151 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
152 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
153 _mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC));
154 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
155
156 a = (const void**restrict) ((uintptr_t) a - ks);
157 nc -= 16;
158 } else {
159 __m128i vh3x01234567 = _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC);
160 __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
161 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
162 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
163 if (nc & 8) {
164 _mm_storeu_si128((__m128i*) c3, vh3x01234567);
165 _mm_storeu_si128((__m128i*) c2, vh2x01234567);
166 _mm_storeu_si128((__m128i*) c1, vh1x01234567);
167 _mm_storeu_si128((__m128i*) c0, vh0x01234567);
168
169 vh3x01234567 = _mm256_cvtps_ph(vacc3x89ABCDEF, _MM_FROUND_NO_EXC);
170 vh2x01234567 = _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC);
171 vh1x01234567 = _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC);
172 vh0x01234567 = _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC);
173
174 c3 += 8;
175 c2 += 8;
176 c1 += 8;
177 c0 += 8;
178 }
179 if (nc & 4) {
180 _mm_storel_epi64((__m128i*) c3, vh3x01234567);
181 _mm_storel_epi64((__m128i*) c2, vh2x01234567);
182 _mm_storel_epi64((__m128i*) c1, vh1x01234567);
183 _mm_storel_epi64((__m128i*) c0, vh0x01234567);
184
185 vh3x01234567 = _mm_unpackhi_epi64(vh3x01234567, vh3x01234567);
186 vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
187 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
188 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
189
190 c3 += 4;
191 c2 += 4;
192 c1 += 4;
193 c0 += 4;
194 }
195 if (nc & 2) {
196 _mm_storeu_si32(c3, vh3x01234567);
197 _mm_storeu_si32(c2, vh2x01234567);
198 _mm_storeu_si32(c1, vh1x01234567);
199 _mm_storeu_si32(c0, vh0x01234567);
200
201 vh3x01234567 = _mm_srli_epi64(vh3x01234567, 32);
202 vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
203 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
204 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
205
206 c3 += 2;
207 c2 += 2;
208 c1 += 2;
209 c0 += 2;
210 }
211 if (nc & 1) {
212 *c3 = _mm_extract_epi16(vh3x01234567, 0);
213 *c2 = _mm_extract_epi16(vh2x01234567, 0);
214 *c1 = _mm_extract_epi16(vh1x01234567, 0);
215 *c0 = _mm_extract_epi16(vh0x01234567, 0);
216 }
217
218 nc = 0;
219 }
220 } while (nc != 0);
221 }
222