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