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_3x16__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_3x16__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 <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint16_t) == 0);
37 assert(ks != 0);
38 assert(ks % (3 * 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
54 do {
55 __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
56 __m256 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
57 __m256 vacc1x01234567 = vacc0x01234567;
58 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
59 __m256 vacc2x01234567 = vacc0x01234567;
60 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
61 w = (const uint16_t*) w + 16;
62
63 size_t p = ks;
64 do {
65 const uint16_t* restrict a0 = (const uint16_t*) a[0];
66 assert(a0 != NULL);
67 if XNN_UNPREDICTABLE(a0 != zero) {
68 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
69 }
70 const uint16_t* restrict a1 = (const uint16_t*) a[1];
71 assert(a1 != NULL);
72 if XNN_UNPREDICTABLE(a1 != zero) {
73 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
74 }
75 const uint16_t* restrict a2 = (const uint16_t*) a[2];
76 assert(a2 != NULL);
77 if XNN_UNPREDICTABLE(a2 != zero) {
78 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
79 }
80 a += 3;
81
82 size_t k = kc;
83 do {
84 const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
85 const __m256 vb89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
86 w = (const uint16_t*) w + 16;
87
88 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
89 a0 += 1;
90 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
91 a1 += 1;
92 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
93 a2 += 1;
94
95 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
96 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF), _MM_FROUND_NO_EXC));
97 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
98 vacc1x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF), _MM_FROUND_NO_EXC));
99 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
100 vacc2x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF), _MM_FROUND_NO_EXC));
101
102 k -= sizeof(uint16_t);
103 } while (k != 0);
104 p -= 3 * sizeof(void*);
105 } while (p != 0);
106
107 const __m256 vmin = _mm256_load_ps(params->avx.min);
108 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
109 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
110 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
111 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
112 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
113 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
114
115 const __m256 vmax = _mm256_load_ps(params->avx.max);
116 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
117 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
118 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
119 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
120 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
121 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
122
123 if XNN_LIKELY(nc >= 16) {
124 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
125 _mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC));
126 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
127 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
128 _mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC));
129 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
130 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
131 _mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC));
132 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
133
134 a = (const void**restrict) ((uintptr_t) a - ks);
135 nc -= 16;
136 } else {
137 __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
138 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
139 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
140 if (nc & 8) {
141 _mm_storeu_si128((__m128i*) c2, vh2x01234567);
142 _mm_storeu_si128((__m128i*) c1, vh1x01234567);
143 _mm_storeu_si128((__m128i*) c0, vh0x01234567);
144
145 vh2x01234567 = _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC);
146 vh1x01234567 = _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC);
147 vh0x01234567 = _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC);
148
149 c2 += 8;
150 c1 += 8;
151 c0 += 8;
152 }
153 if (nc & 4) {
154 _mm_storel_epi64((__m128i*) c2, vh2x01234567);
155 _mm_storel_epi64((__m128i*) c1, vh1x01234567);
156 _mm_storel_epi64((__m128i*) c0, vh0x01234567);
157
158 vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
159 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
160 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
161
162 c2 += 4;
163 c1 += 4;
164 c0 += 4;
165 }
166 if (nc & 2) {
167 _mm_storeu_si32(c2, vh2x01234567);
168 _mm_storeu_si32(c1, vh1x01234567);
169 _mm_storeu_si32(c0, vh0x01234567);
170
171 vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
172 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
173 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
174
175 c2 += 2;
176 c1 += 2;
177 c0 += 2;
178 }
179 if (nc & 1) {
180 *c2 = _mm_extract_epi16(vh2x01234567, 0);
181 *c1 = _mm_extract_epi16(vh1x01234567, 0);
182 *c0 = _mm_extract_epi16(vh0x01234567, 0);
183 }
184
185 nc = 0;
186 }
187 } while (nc != 0);
188 }
189