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_5x8__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_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_5x8__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_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 5);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint16_t) == 0);
37 assert(ks != 0);
38 assert(ks % (5 * 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 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61
62 do {
63 __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
64 __m256 vacc1x01234567 = vacc0x01234567;
65 __m256 vacc2x01234567 = vacc0x01234567;
66 __m256 vacc3x01234567 = vacc0x01234567;
67 __m256 vacc4x01234567 = vacc0x01234567;
68 w = (const uint16_t*) w + 8;
69
70 size_t p = ks;
71 do {
72 const uint16_t* restrict a0 = (const uint16_t*) a[0];
73 assert(a0 != NULL);
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
76 }
77 const uint16_t* restrict a1 = (const uint16_t*) a[1];
78 assert(a1 != NULL);
79 if XNN_UNPREDICTABLE(a1 != zero) {
80 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
81 }
82 const uint16_t* restrict a2 = (const uint16_t*) a[2];
83 assert(a2 != NULL);
84 if XNN_UNPREDICTABLE(a2 != zero) {
85 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
86 }
87 const uint16_t* restrict a3 = (const uint16_t*) a[3];
88 assert(a3 != NULL);
89 if XNN_UNPREDICTABLE(a3 != zero) {
90 a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset);
91 }
92 const uint16_t* restrict a4 = (const uint16_t*) a[4];
93 assert(a4 != NULL);
94 if XNN_UNPREDICTABLE(a4 != zero) {
95 a4 = (const uint16_t*) ((uintptr_t) a4 + a_offset);
96 }
97 a += 5;
98
99 size_t k = kc;
100 do {
101 const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
102 w = (const uint16_t*) w + 8;
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 const __m256 va4 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a4));
113 a4 += 1;
114
115 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
116 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
117 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
118 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb01234567, vacc3x01234567), _MM_FROUND_NO_EXC));
119 vacc4x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va4, vb01234567, vacc4x01234567), _MM_FROUND_NO_EXC));
120
121 k -= sizeof(uint16_t);
122 } while (k != 0);
123 p -= 5 * sizeof(void*);
124 } while (p != 0);
125
126 const __m256 vscale = _mm256_load_ps(params->avx.scale);
127 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc0x01234567, vscale), _MM_FROUND_NO_EXC));
128 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc1x01234567, vscale), _MM_FROUND_NO_EXC));
129 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc2x01234567, vscale), _MM_FROUND_NO_EXC));
130 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc3x01234567, vscale), _MM_FROUND_NO_EXC));
131 vacc4x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc4x01234567, vscale), _MM_FROUND_NO_EXC));
132
133 const __m256 vmin = _mm256_load_ps(params->avx.min);
134 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
135 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
136 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
137 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
138 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
139
140 const __m256 vmax = _mm256_load_ps(params->avx.max);
141 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
142 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
143 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
144 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
145 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
146
147 if XNN_LIKELY(nc >= 8) {
148 _mm_storeu_si128((__m128i*) c4, _mm256_cvtps_ph(vacc4x01234567, _MM_FROUND_NO_EXC));
149 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
150 _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC));
151 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
152 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
153 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
154 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
155 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
156 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
157 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
158
159 a = (const void**restrict) ((uintptr_t) a - ks);
160 nc -= 8;
161 } else {
162 __m128i vh4x01234567 = _mm256_cvtps_ph(vacc4x01234567, _MM_FROUND_NO_EXC);
163 __m128i vh3x01234567 = _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC);
164 __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
165 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
166 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
167 if (nc & 4) {
168 _mm_storel_epi64((__m128i*) c4, vh4x01234567);
169 _mm_storel_epi64((__m128i*) c3, vh3x01234567);
170 _mm_storel_epi64((__m128i*) c2, vh2x01234567);
171 _mm_storel_epi64((__m128i*) c1, vh1x01234567);
172 _mm_storel_epi64((__m128i*) c0, vh0x01234567);
173
174 vh4x01234567 = _mm_unpackhi_epi64(vh4x01234567, vh4x01234567);
175 vh3x01234567 = _mm_unpackhi_epi64(vh3x01234567, vh3x01234567);
176 vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
177 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
178 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
179
180 c4 += 4;
181 c3 += 4;
182 c2 += 4;
183 c1 += 4;
184 c0 += 4;
185 }
186 if (nc & 2) {
187 _mm_storeu_si32(c4, vh4x01234567);
188 _mm_storeu_si32(c3, vh3x01234567);
189 _mm_storeu_si32(c2, vh2x01234567);
190 _mm_storeu_si32(c1, vh1x01234567);
191 _mm_storeu_si32(c0, vh0x01234567);
192
193 vh4x01234567 = _mm_srli_epi64(vh4x01234567, 32);
194 vh3x01234567 = _mm_srli_epi64(vh3x01234567, 32);
195 vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
196 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
197 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
198
199 c4 += 2;
200 c3 += 2;
201 c2 += 2;
202 c1 += 2;
203 c0 += 2;
204 }
205 if (nc & 1) {
206 *c4 = _mm_extract_epi16(vh4x01234567, 0);
207 *c3 = _mm_extract_epi16(vh3x01234567, 0);
208 *c2 = _mm_extract_epi16(vh2x01234567, 0);
209 *c1 = _mm_extract_epi16(vh1x01234567, 0);
210 *c0 = _mm_extract_epi16(vh0x01234567, 0);
211 }
212
213 nc = 0;
214 }
215 } while (nc != 0);
216 }
217