1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gemm/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/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_f16_gemm_minmax_ukernel_3x16__avx2_broadcast(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_gemm_minmax_ukernel_3x16__avx2_broadcast(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const void*restrict a,
23 size_t a_stride,
24 const void*restrict w,
25 void*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint16_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const uint16_t* a0 = a;
40 uint16_t* c0 = c;
41 const uint16_t* a1 = (const uint16_t*) ((uintptr_t) a0 + a_stride);
42 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const uint16_t* a2 = (const uint16_t*) ((uintptr_t) a1 + a_stride);
48 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
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 k = kc;
64 do {
65 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
66 a0 += 1;
67 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
68 a1 += 1;
69 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
70 a2 += 1;
71
72 const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
73 const __m256 vb89ABCDEF = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8)));
74 w = (const uint16_t*) w + 16;
75
76 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
77 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
78 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
79 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF), _MM_FROUND_NO_EXC));
80 vacc1x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF), _MM_FROUND_NO_EXC));
81 vacc2x89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF), _MM_FROUND_NO_EXC));
82
83 k -= sizeof(uint16_t);
84 } while (k != 0);
85
86 const __m256 vmin = _mm256_load_ps(params->avx.min);
87 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
88 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
89 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
90 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
91 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
92 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
93
94 const __m256 vmax = _mm256_load_ps(params->avx.max);
95 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
96 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
97 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
98 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
99 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
100 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
101
102 if XNN_LIKELY(nc >= 16) {
103 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
104 _mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC));
105 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
106 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
107 _mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC));
108 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
109 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
110 _mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC));
111 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
112
113 a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
114 a1 = (const uint16_t*) ((uintptr_t) a1 - kc);
115 a2 = (const uint16_t*) ((uintptr_t) a2 - kc);
116
117 nc -= 16;
118 } else {
119 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
120 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
121 __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
122 if (nc & 8) {
123 _mm_storeu_si128((__m128i*) c0, vh0x01234567);
124 _mm_storeu_si128((__m128i*) c1, vh1x01234567);
125 _mm_storeu_si128((__m128i*) c2, vh2x01234567);
126
127 vh0x01234567 = _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC);
128 vh1x01234567 = _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC);
129 vh2x01234567 = _mm256_cvtps_ph(vacc2x89ABCDEF, _MM_FROUND_NO_EXC);
130
131 c0 += 8;
132 c1 += 8;
133 c2 += 8;
134 }
135 if (nc & 4) {
136 _mm_storel_epi64((__m128i*) c0, vh0x01234567);
137 _mm_storel_epi64((__m128i*) c1, vh1x01234567);
138 _mm_storel_epi64((__m128i*) c2, vh2x01234567);
139
140 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
141 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
142 vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
143
144 c0 += 4;
145 c1 += 4;
146 c2 += 4;
147 }
148 if (nc & 2) {
149 _mm_storeu_si32(c0, vh0x01234567);
150 _mm_storeu_si32(c1, vh1x01234567);
151 _mm_storeu_si32(c2, vh2x01234567);
152
153 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
154 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
155 vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
156
157 c0 += 2;
158 c1 += 2;
159 c2 += 2;
160 }
161 if (nc & 1) {
162 *c0 = (uint16_t) _mm_extract_epi16(vh0x01234567, 0);
163 *c1 = (uint16_t) _mm_extract_epi16(vh1x01234567, 0);
164 *c2 = (uint16_t) _mm_extract_epi16(vh2x01234567, 0);
165 }
166
167 nc = 0;
168 }
169 } while (nc != 0);
170 }
171