1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/avx-broadcast.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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
16
xnn_f32_igemm_minmax_ukernel_3x16__fma3_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_3x16__fma3_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52
53 do {
54 __m256 vacc0x01234567 = _mm256_load_ps(w);
55 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
56 __m256 vacc1x01234567 = vacc0x01234567;
57 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
58 __m256 vacc2x01234567 = vacc0x01234567;
59 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
60 w += 16;
61
62 size_t p = ks;
63 do {
64 const float* restrict a0 = a[0];
65 assert(a0 != NULL);
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const float*) ((uintptr_t) a0 + a_offset);
68 }
69 const float* restrict a1 = a[1];
70 assert(a1 != NULL);
71 if XNN_UNPREDICTABLE(a1 != zero) {
72 a1 = (const float*) ((uintptr_t) a1 + a_offset);
73 }
74 const float* restrict a2 = a[2];
75 assert(a2 != NULL);
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const float*) ((uintptr_t) a2 + a_offset);
78 }
79 a += 3;
80
81 size_t k = kc;
82 do {
83 const __m256 vb01234567 = _mm256_load_ps(w);
84 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
85 w += 16;
86
87 const __m256 va0 = _mm256_broadcast_ss(a0);
88 a0 += 1;
89 const __m256 va1 = _mm256_broadcast_ss(a1);
90 a1 += 1;
91 const __m256 va2 = _mm256_broadcast_ss(a2);
92 a2 += 1;
93
94 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
95 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
96 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
97 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF);
98 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
99 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);
100 k -= sizeof(float);
101 } while (k != 0);
102 p -= 3 * sizeof(void*);
103 } while (p != 0);
104
105 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
106 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
107 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
108 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
109 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
110 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
111 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
112
113 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
114 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
115 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
116 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
117 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
118 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
119 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
120
121 if XNN_LIKELY(nc >= 16) {
122 _mm256_storeu_ps(c2, vacc2x01234567);
123 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
124 c2 = (float*) ((uintptr_t) c2 + cn_stride);
125 _mm256_storeu_ps(c1, vacc1x01234567);
126 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
127 c1 = (float*) ((uintptr_t) c1 + cn_stride);
128 _mm256_storeu_ps(c0, vacc0x01234567);
129 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
130 c0 = (float*) ((uintptr_t) c0 + cn_stride);
131
132 a = (const float**restrict) ((uintptr_t) a - ks);
133 nc -= 16;
134 } else {
135 if (nc & 8) {
136 _mm256_storeu_ps(c2, vacc2x01234567);
137 _mm256_storeu_ps(c1, vacc1x01234567);
138 _mm256_storeu_ps(c0, vacc0x01234567);
139
140 vacc2x01234567 = vacc2x89ABCDEF;
141 vacc1x01234567 = vacc1x89ABCDEF;
142 vacc0x01234567 = vacc0x89ABCDEF;
143
144 c2 += 8;
145 c1 += 8;
146 c0 += 8;
147 }
148 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
149 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
150 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
151 if (nc & 4) {
152 _mm_storeu_ps(c2, vacc2x0123);
153 _mm_storeu_ps(c1, vacc1x0123);
154 _mm_storeu_ps(c0, vacc0x0123);
155
156 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
157 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
158 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
159
160 c2 += 4;
161 c1 += 4;
162 c0 += 4;
163 }
164 if (nc & 2) {
165 _mm_storel_pi((__m64*) c2, vacc2x0123);
166 _mm_storel_pi((__m64*) c1, vacc1x0123);
167 _mm_storel_pi((__m64*) c0, vacc0x0123);
168
169 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
170 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
171 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
172
173 c2 += 2;
174 c1 += 2;
175 c0 += 2;
176 }
177 if (nc & 1) {
178 _mm_store_ss(c2, vacc2x0123);
179 _mm_store_ss(c1, vacc1x0123);
180 _mm_store_ss(c0, vacc0x0123);
181 }
182
183 nc = 0;
184 }
185 } while (nc != 0);
186 }
187